Paper deep dive
Cross-Modal Navigation with Multi-Agent Reinforcement Learning
Shuo Liu, Xinzichen Li, Christopher Amato
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 7/8/2026, 11:03:18 AM
Summary
The paper introduces CRONA, a Multi-Agent Reinforcement Learning framework for cross-modal navigation. It addresses multi-modal data alignment and policy complexity by using lightweight, modality-specialized agents that collaborate decentralizedly. CRONA leverages auxiliary belief predictors, an attention-based history encoder, and a centralized critic with global state information. Experiments on visual-acoustic tasks show that multi-agent collaboration significantly outperforms single-agent baselines, with heterogeneous collaboration proving most effective for complex environments.
Entities (10)
Relation Signals (10)
CRONA → uses → Multi-Agent Reinforcement Learning
confidence 97% · In this paper, we propose CRONA, a Multi-Agent Reinforcement Learning (MARL) framework for Cross-Modal Navigation.
CRONA → targets → Visual-Acoustic Navigation
confidence 95% · Experiments on visual-acoustic navigation tasks show that multi-agent methods significantly improve performance
CRONA → employs → Auxiliary Belief Predictor
confidence 94% · CRONA employs auxiliary belief predictors to extract control-relevant features from complex multi-modal observations
CRONA → employs → Centralized Critic
confidence 94% · CRONA employs a centralized critic for joint value estimation during training.
CRONA → evaluatedon → Matterport3D
confidence 93% · We construct a multi-modal collaborative navigation benchmark based on diverse Matterport3D scenes
Centralized Critic → utilizes → Global State
confidence 93% · The centralized critic estimates the joint value from the joint history, beliefs, and global state during training
CRONA → employs → Attention-Based History Encoder
confidence 92% · We use a short-term history cache and apply multi-head attention to extract spatial and temporal features.
Heterogeneous Collaboration → →
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Robust embodied navigation relies on complementary sensory cues. However, high-quality and well-aligned multi-modal data is often difficult to obtain in practice. Training a monolithic model is also challenging as rich multi-modal inputs induce complex representations and substantially enlarge the policy space. Cross-modal collaboration among lightweight modality-specialized agents offers a scalable paradigm. It enables flexible deployment and parallel execution, while preserving the strength of each modality. In this paper, we propose \textbf{CRONA}, a Multi-Agent Reinforcement Learning (MARL) framework for \textbf{Cro}ss-Modal \textbf{Na}vigation. CRONA improves collaboration by leveraging control-relevant auxiliary beliefs and a centralized multi-modal critic with global state. Experiments on visual-acoustic navigation tasks show that multi-agent methods significantly improve performance and efficiency over single-agent baselines. We find that homogeneous collaboration with limited modalities is sufficient for short-range navigation under salient cues; heterogeneous collaboration among agents with complementary modalities is generally efficient and effective; and navigation in large, complex environments requires both richer multi-modal perception and increased model capacity.
Tags
Links
- Source: https://arxiv.org/abs/2605.06595v1
- Canonical: https://arxiv.org/abs/2605.06595v1
Trouble viewing inline? Open PDF directly →
Full Text
82,315 characters extracted from source content.
Expand or collapse full text
Cross-Modal Navigation with Multi-Agent Reinforcement Learning Shuo Liu, Xinzichen Li, Christopher Amato Khoury College of Computer Sciences Northeastern University Boston, MA, 02120 liu.shuo2,li.xinzi,c.amato@northeastern.edu Abstract Robust embodied navigation relies on complementary sensory cues. However, high-quality and well-aligned multi-modal data is often difficult to obtain in practice. Training a monolithic model is also challenging as rich multi-modal inputs induce complex representations and substantially enlarge the policy space. Cross-modal collaboration among lightweight modality-specialized agents offers a scalable paradigm. It enables flexible deployment and parallel execution, while preserving the strength of each modality. In this paper, we propose CRONA, a Multi-Agent Reinforcement Learning (MARL) framework for Cross-Modal Navigation. CRONA improves collaboration by leveraging control-relevant auxiliary beliefs and a centralized multi-modal critic with global state. Experiments on visual-acoustic navigation tasks show that multi-agent methods significantly improve performance and efficiency over single-agent baselines. We find that homogeneous collaboration with limited modalities is sufficient for short-range navigation under salient cues; heterogeneous collaboration among agents with complementary modalities is generally efficient and effective; and navigation in large, complex environments requires both richer multi-modal perception and increased model capacity. 1 Introduction Figure 1: A collaborative navigation task in a Ranch scene from Matterport3D. An audio agent (blue) collaborates with a vision agent (green) to locate a table and pictures. Each agent receives only local observations during execution, while global information is captured by a global monitor (yellow) and used only during training. Gray curves denote agents’ trajectories. In embodied navigation, agents perceive the environment through diverse sensory inputs, e.g., RGB-D images, audio, radar, LiDAR, and language instructions Kolve et al. (2017); Manolis Savva* et al. (2019); Szot et al. (2021); Anderson et al. (2018); Duan et al. (2022). These inputs provide rich geometric, semantic, and acoustic cues across modalities, enabling agents to locate target objects and navigate in complex environments, such as autonomous driving, robotic systems, and human-computer interaction Chen et al. (2017); Das et al. (2017); Ku et al. (2018); Qi et al. (2018); Shridhar et al. (2022); Brohan et al. (2022); Driess et al. (2023). However, real-world observations are often noisy, incomplete, and asynchronous. Low-quality and misaligned training signals can make policy learning unstable and ineffective Tsai et al. (2019); Jia et al. (2021). Although many methods align different modalities within the model, they differ substantially in dimensionality, noise levels, and temporal structure Chen et al. (2020a); Shridhar et al. (2020b, a). This mismatch often leads to imbalanced joint optimization, where dominant modalities drive most gradient updates while weaker or noisier modalities are underutilized Wang et al. (2020); Wu et al. (2022); Huang et al. (2022); Peng et al. (2022). Rich-modal models also tend to rely on large architectures to align diverse signals, making them hard to deploy and costly at test time Kim et al. (2021); Yao et al. (2024); Li et al. (2023). Many studies leverage multi-agent collaboration to improve navigation efficiency and robustness Simmons et al. (2000); Parker (2002); Burgard et al. (2005); Gu et al. (2021). However, most existing approaches focus on collaboration under limited sensory configurations Huang et al. (2021); Qin et al. (2021); Azzam et al. (2023); Xiao et al. (2020, 2022); Xue and Chen (2023); Wang et al. (2024). Even in more complex embodied settings, collaborative agents are typically homogeneous, with each receiving inputs from the same modalities Wang et al. (2021); Zhang et al. (2025); Wang et al. (2026). Heterogeneous collaboration with rich sensory modalities are less explored Hao et al. (2025); Liu et al. (2025); Hu et al. (2023). Specifically, it remains unclear which modalities improve collaborative capability, which team configurations support both effective and efficient navigation, and what cooperative behaviors emerge among agents. In this paper, we study fully-decentralized collaborative navigation without any inter-agent communication Oliehoek et al. (2008); Oliehoek and Amato (2016). We construct a multi-modal collaborative navigation benchmark based on diverse Matterport3D scenes, as illustrated in Figure 1. We propose CRONA, a cooperative Multi-Agent Reinforcement Learning (MARL) framework for Cross-Modal Navigation. CRONA employs auxiliary belief predictors to extract control-relevant features from complex multi-modal observations and a centralized critic with state information to facilitate training. Our experiments demonstrate that collaborative navigation consistently outperforms single-agent navigation in both effectiveness and efficiency. Moreover, we identify five modality-dominance patterns across scenarios (i.e., no clear dominance, vision dominance, audio dominance, cross-modal, and multi-modal dominance). We find that homogeneous collaboration with few modalities is sufficient for short-range navigation; cross-modal collaboration among complementary modalities is generally efficient and effective when targets have clean, modality-specific cues; large and complex environments typically require both full-modal inputs and higher-capacity models. Our core contributions are summarized as follows: (i) we construct a collaborative navigation benchmark where multi-modal agents collaborate to navigate; (i) we propose CRONA, a MARL framework for cross-modal navigation; (i) we identify five dominance patterns in our experiments and explain when and why they emerge, respectively. 2 Related Work Multi-Modal Navigation Embodied navigation has been studied under a wide range of input modalities. Most work considers visual observations (RGB-D images), while specifying navigation goals or instructions in language Anderson et al. (2018); Krantz et al. (2020); Fried et al. (2018); Hao et al. (2020); Majumdar et al. (2020); Hong et al. (2021); Chen et al. (2021b). Acoustic and 3D spatial signals can also provide semantic and geometric cues that complement visual observations that are degraded by occlusions or obstacles Cheng and Wang (2018); Chen et al. (2020a, 2021a); Paul et al. (2022); Chen et al. (2020b); Yao et al. (2023); Yuksel et al. (2026). While it has been shown that richer multi-modal context can improve performance in certain settings Qi et al. (2021); Huang et al. (2023); Yu et al. (2023); Liu et al. (2025), it remains unclear how different modalities contribute under different conditions and how to align modalities with substantially different representations Wang et al. (2020); Wu et al. (2022); Huang et al. (2022); Li et al. (2023); Peng et al. (2022). Collaborative Navigation Recent studies have explored multi-agent collaboration for navigation. However, collaborative navigation remains challenging because agents need to coordinate in real time during execution. Early methods rely on centralized planning, where a central controller coordinates all agents Bruce and Veloso (2006); van Den Berg et al. (2009); Janssen et al. (2016). Such designs suffer from limited scalability and robustness Velagapudi et al. (2010); Luna and Bekris (2011); Iqbal and Sha (2019). Recent approaches therefore shift toward decentralized collaboration, where agents take action based on their local observations with limited or even without communication Huang et al. (2021); Qin et al. (2021); Azzam et al. (2023); Wang et al. (2024); Xue and Chen (2023); Wang et al. (2026). However, most decentralized navigation studies still assume homogeneous agents or agents with similar sensory inputs Zhang et al. (2025); Xiao et al. (2020, 2022). How agents with heterogeneous modalities collaborate effectively remains largely underexplored Hao et al. (2025); Liu et al. (2025); Hu et al. (2023). Cooperative MARL Cooperative MARL studies how multiple agents learn to coordinate under a shared objective Zhang et al. (2021); Albrecht et al. (2024); Yuan et al. (2023). A simple and scalable approach is independent learning, where agents are separately trained Tan and others (1993); Peshkin et al. (2001). But as all agents update their policies concurrently, each agent faces a non-stationary learning environment, which often leads to instability and convergence issues Claus and Boutilier (1998); Tuyls et al. (2003); Wunder et al. (2010). Centralized training with decentralized execution (CTDE) mitigates this issue by exploiting centralized information during training Amato (2024). For example, a centralized critic can estimate joint values from joint histories and global states Lowe et al. (2017); Yu et al. (2022); Foerster et al. (2018); Lyu et al. (2021, 2023). Since the critic is discarded at execution time, each agent remains execute in a decentralized manner. CRONA follows this paradigm and incorporates task progress into a multi-modal centralized critic for joint value estimation. 3 Background 3.1 Problem Formulation In cooperative navigation (Figure 1), agents need to infer task assignments and learn cooperative policies under partial observations. This setting follows the standard cooperative MARL formulation and can be modeled as a Decentralized Partially Observable Markov Decision Process (Dec-POMDP) Oliehoek and Amato (2016), denoted by ⟨ℐ,,i,i,R,T,γ,H⟩ ,S,\O_i\,\A_i\,R,T,γ,H . • ℐI is a set of n decentralized agents, where each agent i is controlled by an individual policy πi _i. Each agent is equipped with specialized sensors to perceive the environment. • S denotes the global state space. At each time step t, the global state st∈s_t includes the scene layout, all agent poses, target object locations and categories, sound-source states, and task-completion status. This state is not directly observed by decentralized agents. • Each agent i receives a local observation oi,t∈io_i,t _i. The observation contains the agent pose oi,tpose=(xi,t,yi,t,ϑi,t,t)o_i,t^pose=(x_i,t,y_i,t, _i,t,t), where (xi,t,yi,t)(x_i,t,y_i,t) is the agent position and ϑi,t _i,t is its orientation. It also includes a natural-language description of the navigation target, denoted by oi,tgoalo_i,t^goal. Depending on its sensor configuration, an agent may also receive visual input oi,tvision=(oi,trgb,oi,tdepth)∈ℝHv×Wv×4o_i,t^vision=(o_i,t^rgb,o_i,t^depth) ^H_v× W_v× 4, binaural audio input oi,taudio∈ℝ2×Lo_i,t^audio ^2× L, where HvH_v and WvW_v denote the image height and width, and L denotes the length of the binaural audio segment. iO_i is the local observation space of agent i, and =×iO=×_iO_i is the joint observation space. • Agents share a joint reward function R:×→ℝR:S×A , which depends on the global state and their joint action. The reward incentives agents to approach targets and stop in their vicinity. • The environment evolves according to a transition function T:×→Δ()T:S×A→ (S). Given the current state sts_t and joint action ta_t, the next state is sampled as st+1∼T(⋅∣st,t)s_t+1 T(· s_t,a_t). • γ is the discount factor and H is the episode horizon. Since the full state is not directly observable, each agent maintains a local observation-action history hi,t=oi,0,ai,0,⋯,oi,th_i,t=\o_i,0,a_i,0,·s,o_i,t\ to infer information about sts_t. The history of agents forms a joint history t=h1,t,⋯,hn,th_t=\h_1,t,·s,h_n,t\, and agents’ policies forms a joint policy =π1,⋯,πn π=\ _1,·s, _n\. The objective is to find an optimal joint policy, ∗=π1∗,⋯,πn∗ π^*=\π^*_1,·s,π^*_n\, that maximizes the expected cumulative reward over the horizon H, ∗=argmax[∑t=0H−1γtrt] π^*= _ πE_ π [ _t=0^H-1γ^tr_t ]. 3.2 Acoustic Representation Audio signals provide semantic information for source recognition and spatial cues for source localization. However, raw audio waves are high-dimensional and contain complex temporal dependencies. These make them difficult to model directly. Spectrograms represent audio as structured time-frequency features, making local acoustic patterns more explicit and easier to learn. Given a binaural waveform oi,taudio∈ℝ2×Lo_i,t^audio ^2× L, its magnitude spectrogram oi,tspec∈ℝ2×K×Fo_i,t^spec ^2× K× F can be computed via short-time Fourier transform (STFT), oi,tspec(κ,ω,τ)=|∑ℓ=0L−1oi,taudio(κ,ℓ)w(ℓ−τδ)e−j2πωℓ/Nfft|.o_i,t^spec(κ,ω,τ)= | _ =0^L-1o_i,t^audio(κ, )\,w( -τδ)\,e^-j2πω /N_fft |. (1) Here, κ∈1,2κ∈\1,2\ denotes the left and right audio channels. For each time frame τ, the window w(⋅)w(·) extracts a short segment of the waveform around sample position τδτδ. The Fourier basis then decomposes this segment into frequency components indexed by ω. δ denotes the hop size between adjacent time frames, and NfftN_fft denotes the FFT size. The resulting spectrogram contains K frequency bins over F time frames. Sounds emitted by different objects produce distinct patterns in the spectrogram. Targets with salient acoustic cues, such as strong energy and clean, stable patterns, are typically easier to localize, whereas distant or occluded sounds tend to be weak, unclear, and difficult to identify. 4 Method Figure 2 gives an overview of CRONA. Each agent processes its sensory observations with the corresponding encoder. Audio-based agents use an auxiliary belief predictor to estimate control-relevant beliefs (target category and location). Each agent combines its observations, beliefs, and previous actions into a local history, where multi-head attention layers capture important features and temporal dependencies. A centralized critic estimates the joint value from the joint history, beliefs, and global state during training, which is used to update decentralized agent policies. 4.1 Auxiliary Belief Predictor Audio observations are often noisy and stochastic (Section 3.2), making it difficult to learn effective policies directly from raw inputs. However, control-relevant beliefs can be inferred from these signals to facilitate training. CRONA uses target location and target category as auxiliary beliefs. For an agent i with audio sensor, given its spectrogram observation oi,tspeco_i,t^spec, a convolutional encoder extracts acoustic features zi,taudioz_i,t^audio. A location head predicts an instantaneous sound-source goal b^i,tgoal∈ℝ2 b_i,t^goal ^2 in global coordinate based on zi,taudioz_i,t^audio. Given the current pose oi,tpose=(xi,t,yi,t,ϑi,t,t)o_i,t^pose=(x_i,t,y_i,t, _i,t,t), the predicted relative location b^i,tloc b_i,t^loc can be calculated by, b^i,tloc=T(ϑi,t)(b^i,tgoal−[xi,tyi,t]), b_i,t^loc=T( _i,t) ( b_i,t^goal- bmatrixx_i,t\\ y_i,t bmatrix ), (2) where T(ϑ)=[cosϑsinϑ−sinϑcosϑ]T( )= bmatrix & \\ - & bmatrix is a 2D rotation matrix from the global frame to the agent’s frame. In addition, a category head with fully connected layers also predicts a belief b^i,tcat∈ℝ\, b_i,t^cat ^C over all categories c∈c based on zi,taudioz_i,t^audio. To reduce prediction variance, we smooth the auxiliary beliefs with an exponential moving average using coefficient α∈[0,1]α∈[0,1], bi,tloc=αb^i,tloc+(1−α)bi,t−1loc,bi,tcat=αb^i,tcat+(1−α)bi,t−1cat.b_i,t^loc=α\, b_i,t^loc+(1-α)\,b_i,t-1^loc, b_i,t^cat=α\, b_i,t^cat+(1-α)\,b_i,t-1^cat. (3) The location and category belief jointly form an auxiliary belief bi,t=(bi,tloc,bi,tcat)b_i,t=(b_i,t^loc,b_i,t^cat) for agent i, and since they are inferred from local histories, they remain consistent with the information available to decentralized policies. During training, the goal point bi,tgoal,∗b_i,t^goal,* of the closest target to agent i and the multi-hot category label yi,tcat,∗y_i,t^cat,* over all targets are provided. The belief predictor is optimized as, ℒbelief=‖b^i,tgoal−bi,tgoal,∗‖22−∑c∈[yi,tcat,∗(c)logb^i,tcat(c)+(1−yi,tcat,∗(c))log(1−b^i,tcat(c))].L_belief= \| b_i,t^goal-b_i,t^goal,* \|_2^2- _c [y_i,t^cat,*(c) b_i,t^cat(c)+ (1-y_i,t^cat,*(c) ) (1- b_i,t^cat(c) ) ]. (4) 4.2 Attention-Based History Encoder In collaborative navigation, each agent selects actions based on its history. However, maintaining all raw images and audio over time is computationally expensive and difficult to optimize. We use a short-term history cache and apply multi-head attention to extract spatial and temporal features. We use convolutional encoders to capture the local patterns of RGB-D images and spectrograms, i.e., zi,trgbz_i,t^rgb, zi,tdepthz_i,t^depth, and zi,taudioz_i,t^audio, respectively. The visual inputs oi,trgbo^rgb_i,t and oi,tdeptho^depth_i,t often have higher dimensionalities and exhibit richer spatial structures, whereas oi,tspeco^spec_i,t are computed over short temporal windows, so we use deeper convolutional neural networks as visual encoders (i.e., ResNet-18 He et al. (2016)). The encoded features are concatenated with the agent pose and the goal instruction to form a latent observation embedding zi,to=zi,trgb⊕zi,tdepth⊕oi,tpose⊕oi,tgoalz_i,t^o=z_i,t^rgb z_i,t^depth o_i,t^pose o_i,t^goal for vision-based agents, and zi,to=zi,taudio⊕oi,tpose⊕oi,tgoalz_i,t^o=z_i,t^audio o_i,t^pose o_i,t^goal for audio-based agents. Each agent stores the current observation embedding, the previous k observation embeddings zi,t−ko,…,zi,t−1o\z_i,t-k^o,…,z_i,t-1^o\, and the previous k actions ai,t−k,…,ai,t−1\a_i,t-k,…,a_i,t-1\ in a fixed-size memory cache. The cached observation-action sequence is then processed by transformer blocks to produce a history representation zi,thz_i,t^h over hi,th_i,t. zi,thz_i,t^h can capture geometric cues, sound information, and motion patterns, and provide a compact context for each agent to select its action, i.e., ai,t∼πi(⋅∣hi,t)a_i,t _i(· h_i,t). Figure 2: Illustration of CRONA framework. 2 decentralized agents, one with audio inputs (blue) and another with vision inputs (green), cooperate to navigate toward a table with silverware-dropping sounds and pictures with camera-shutter sounds. (a) Observation-action history embeddings and auxiliary belief predictors of agents. (b) A multi-modal critic (red) estimates the value with joint history, the auxiliary belief, and the global information, while each agent updates its individual policy. 4.3 Centralized Critic with Global Information CRONA employs a centralized critic for joint value estimation during training. Since the reward directly depends on the state, incorporating state information can improve value estimation without introducing bias Lyu et al. (2023). The critic and agent policies also depend on auxiliary belief predictions (Section 4.1), which are consistent with the information available in local observations. The centralized critic is not used during execution, where all agents take action under decentralized policies Albrecht et al. (2024); Amato (2024). During training, the centralized critic ϕ(t,t,st)V_ φ(z^h_t,b_t,s_t) estimates the joint value using the joint history embedding tz^h_t, the auxiliary beliefs of audio-based agents t=b1,t,…,bn,tb_t=\b_1,t,…,b_n,t\, and the global state sts_t (e.g., target locations, agent positions and orientations, and completion indicators for each target). As proved in Appendix A, augmenting the critic with these history-induced beliefs and the global state does not introduce bias in value estimation. At each time step t, the joint history embedding is obtained by concatenating all agents’ history embeddings, t=⨁i=1nzi,thz^h_t= _i=1^nz_i,t^h. To improve representation learning and accelerate training, CRONA shares the modality-specific encoders, auxiliary belief predictor, and history transformer between the decentralized actors and the centralized critic, while using separate heads for policy and value prediction. To stabilize training, we use clipped surrogate objectives for both policy and value updates. The advantage ^t A_t is computed using generalized advantage estimation (GAE), ^t=∑l=0T−t−1(γλ)l[rt+l+γϕold(t+l+1,t+l+1,st+l+1)−ϕold(t+l,t+l,st+l)], A_t= _l=0^T-t-1(γλ)^l [r_t+l+ _ φ_old(h_t+l+1,b_t+l+1,s_t+l+1)-V_ φ_old(h_t+l,b_t+l,s_t+l) ], (5) and the corresponding return target is ^t=^t+ϕold(t,t,st) R_t= A_t+V_ φ_old(h_t,b_t,s_t). We train the value head of the centralized critic by minimizing a clipped value surrogate objective against the return target ^t R_t, L(ϕ)=t[max((ϕ(t,t,st)−^t)2,(¯ϕ(t,t,st)−^t)2)],¯ϕ(t,t,st)=clip(ϕ(t,t,st),ϕold(t,t,st)−ξ,ϕold(t,t,st)+ξ), gatheredL( φ)=E_t [ ( (V_ φ(h_t,b_t,s_t)- R_t )^2,\, ( V_ φ(h_t,b_t,s_t)- R_t )^2 ) ],\\ V_ φ(h_t,b_t,s_t)=clip (V_ φ(h_t,b_t,s_t),V_ φ_old(h_t,b_t,s_t)-ξ,\,V_ φ_old(h_t,b_t,s_t)+ξ ), gathered (6) where ¯ϕ V_ φ denotes the clipped value prediction and ξ is the value clipping range. Each agent’s policy is conditioned only on its local history and auxiliary belief, ai,t∼πθi(⋅∣hi,t,bi,t)a_i,t _ _i(· h_i,t,b_i,t), and is updated using the shared advantage estimate ^t A_t. Specifically, each agent i maximizes J(θi)=t[min(ρi,t^t,clip(ρi,t,1−ϵ,1+ϵ)^t)+βℋ(πθi(⋅∣zi,th,bi,t))],J( _i)=E_t [ ( _i,t A_t,\,clip( _i,t,1-ε,1+ε) A_t )+β\,H\! ( _ _i(· z_i,t^h,b_i,t) ) ], (7) where ρi,t=πθi(ai,t∣zi,th,bi,t)πθi,old(ai,t∣zi,th,bi,t) _i,t= _ _i(a_i,t z_i,t^h,b_i,t) _ _i,old(a_i,t z_i,t^h,b_i,t) is the importance sampling ratio, ϵε is the policy clipping range, β is the entropy regularization coefficient, and ℋ(⋅)H(·) is the policy entropy to encourage exploration. The decentralized actors and the multi-modal critic share encoders and transformers. Gradients from both actor and critic objectives are backpropagated through the shared modules, which are optimized by a weighted sum of the policy gradient and averaged temporal difference loss with μ∈[0,1]μ∈[0,1], ℒ(θizi,ϕzi)=−μJ(θi)+1−μnL(ϕ).L( _i^z_i, φ^z_i)=-μ J( _i)+ 1-μn\,L( φ). (8) 5 Experiments We evaluate CRONA in Matterport3D scenes Chang et al. (2017), where agent observations are simulated via Habitat and libsora Manolis Savva* et al. (2019); Szot et al. (2021); McFee et al. (2015). Dataset details, experimental settings, additional results, instruction and reward designs, and compute resources are provided in Appendix B, D, E, F, and G. 5.1 Setup We construct collaborative navigation datasets with two agents using five representative Matterport3D scenes that span diverse layouts and difficulties. Studio (GdvgFV5R1Z5) is a single-room scene with a picture target with a camera-shutter sound. Corridor (ac26ZMwG7aT) consists of a passage connecting two spatially separated areas, where agents are finding a sink that is dripping water. Apartment (17DRP5sb8fy) has one bedroom and two bathrooms, with a creaking bed and a counter with coin-dropping sound as targets. Ranch (JeFG25nYj2p) contains five bedrooms and two bathrooms, with a picture with a camera-shutter sound and a table with silverware-dropping sounds as targets. Maze (B6ByNegPMKs) is the largest scene with the most complex layout, agents need to find a table with silverware dropping, a dragging chair, and a drawer with a pulling sound while navigating through the scene within the episode limit. Each dataset entry corresponds to a task in an episode. At the beginning of each episode, agents’ positions and orientations are randomly initialized. Agents move on the navigable mesh grids to find all targets. They must stop within a specified distance of a target to mark it as found. Each target sound is assigned to an eligible object with the matching semantic category; sounds from multiple targets are mixed and removed once the corresponding target is found. An episode ends when all targets are found or all agents stop simultaneously. We set the horizon to H=H=70, 150, 500, 1000, 1500 for five scenes. Bird’s-eye-view visualizations and dataset statistics are provided in Appendix B. Since most objects in Matterport3D are large and visually distinctive, an agent can effortlessly localize them without requiring collaboration. However, real-world visual perception is often constrained (e.g., darkness, fog, or blind spots). To make the benchmark more challenging, we restrict vision to depth maps with a sensing range of 0–5m5\,m, a resolution of 16×1616× 16 pixels, and an HFoV of 10∘10 . Details about agent configurations and model architectures are provided in Appendix C.2. 5.2 Baselines We consider the Single-Agent baseline, where a large monolithic model takes all available modalities as input Chen et al. (2020a). For a fair comparison, we use the same episode horizon, and the agent’s initial position is randomly selected from existing initial positions in our collaborative navigation dataset. We further compare CRONA with three homogeneous collaboration baselines, where all agents receive the same input modalities. Several recent studies have explored Vision-Language-Action (VLA) models for collaborative navigation. Hao et al. Hao et al. (2025) propose the CoNav framework in which one agent has access to a bird’s-eye view, while Wang et al. Wang et al. (2026) put forward VLA-based collaborative navigation, CoNavBench, with inter-agent communication. Both settings involve centralized information and differ substantially from ours in environments, agent observability, architectures, and language information. To enable an informative comparison under our task setting, we implement a fully decentralized VLA collaboration baseline as a representative in our scenes, denoted as VLA-Collab. Although audio-language-action (ALA) collaboration has been less explored in navigation, we nevertheless include ALA-Collab as the audio counterpart to VLA-Collab. Both VLA-Collab and ALA-Collab use restricted modality inputs. So we include AVLA-Collab, where all agents receive audio, vision, and language inputs, to represent homogeneous full-modality collaboration in our settings Zhang et al. (2025); Liu et al. (2025). For a fair comparison, all baselines use the same configurations and hyperparameters, and agents in collaborative baselines have the same number of parameters. font=scriptsize (a) Studio ∣ Picture (b) Corridor ∣ Sink (c) Apartment ∣ Bed & Counter (d) Ranch ∣ Picture & Table (e) Maze ∣ Drawer & Table & Chair (f) Aggregated Success Rate Figure 3: Evaluation of CRONA and collaborative navigation baselines across 5 Matterport3D scenes: (a)-(e) show the episode return; (f) shows the aggregated results of success rate. The x-axis indicates the environment steps. Curves are smoothed by an exponential moving average. Shadows denote 90% bootstrapped CI. Results are averaged over 5 runs. 5.3 Results Figure 3 shows the evaluation during training, averaged over five runs. Table 1 provides a detailed comparison between CRONA and the baselines on task completion and navigation efficiency. The effectiveness of collaborative navigation is domain-dependent. We group them into 5 patterns: no clear dominance, vision dominance, audio dominance, cross-modal, and multi-modal dominance. No Clear Dominance As shown in Figure 4(a) and the Studio columns of Table 1, all collaborative navigation methods perform well on Studio, achieving an average success rate of 90.80 ± 4.93%. CRONA achieves the highest success rate, at 95.72%. All collaborative methods substantially outperform the single-agent baseline. This is because decentralized agents can cover a larger exploration area and reduce the impact of premature stopping near the target. These results demonstrate the advantage of collaborative navigation: with proper training, even fully decentralized agents can coordinate effectively without communication. Audio Dominance As shown in Figure 3(b) and the Corridor columns of Table 1, audio cues dominate policy learning in this task. Most methods perform poorly, whereas ALA-Collab achieves the best performance with a 25.31% success rate. This pattern is mainly due to the corridor geometry. Agents initialized near the middle of the corridor receive few informative visual cues and must rely on weak acoustic signals to infer the sound-source direction. As a result, VLA-Collab and AVLA-Collab perform worst among the collaborative baselines, with success rates around 14%. Moreover, incorrect early decisions require long U-turns to recover, as indicated by more than 87.36% timeouts. This audio-dominant pattern suggests that vision is not always the most reliable cue: certain targets can be localized more effectively via audio. This suggests the potential of cross-modal collaboration. Vision Dominance Figure 3(c) and the Apartment columns demonstrate a vision-dominant regime. The collaboration between two vision agents achieves the best performance, reaching a success rate of 78.96%, since the targets are large and visually salient. Audio observations are less reliable in this setting, where two audio agents achieve only 38.23% success, mainly because mixed audio from two distinct sound sources can disrupt auxiliary belief prediction and lead to unstable policy updates. Audio agents struggle to identify the precise stopping location, as reflected by the high early-stop failure rate of 21.45%, compared with 4.96%-11.98% for the other methods. Notably, CRONA outperforms AVLA-Collab by 5.14% in this environment. This suggests that weak or unreliable modalities can hurt multi-modal policies: with limited model capacity, noisy inputs may divert representational capacity away from useful cues. We find that lower target distances and higher detection rates are associated with higher success rates, indicating that task completion is primarily governed by localization quality rather than by a single bottleneck object. Cross-Modal Dominance In Ranch, CRONA achieves the strongest performance, with a 64.62% success rate (Figure 3(d)). We attribute this to effective collaboration between agents with complementary modalities. The audio agent localizes the picture using clean, transient camera-shutter sounds, while the vision agent identifies the table based on its large profile in an open, unobstructed dining room (Appendix B). Interestingly, VLA-Collab and ALA-Collab achieve reasonable performance of around 40%, but AVLA-Collab performs even worse, with only 18.93% success. This is because the monolithic multi-modal with limited capacity struggles to align and effectively exploit different modalities (discussed in Section 5.4). We hypothesize that cross-modal collaboration is particularly effective when different targets have clean, modality-specific cues. It is also parameter-efficient, as each agent only needs to model its own sensory input rather than jointly aligning and reasoning over rich multi-modal observations. Although the success rate is generally consistent with the detection rate and average distance to targets as in other domains, steps, and timeout rate do not align with task success in this scene (Table 1b). Single-Agent, VLA-Collab, and ALA-Collab often terminate early or stop exploring, resulting in fewer steps but lower success. In contrast, AVLA-Collab and CRONA take more exploration steps and achieve higher success. This suggests that in harder, time-constrained tasks, inputs with heterogeneous modalities can induce more diverse behaviors and thereby promote broader exploration. Multi-Modal Dominance In the most complex scene, Maze, collaboration benefits from larger model capacity and access to all available sensory inputs. As shown in Figure 3(e) and Maze columns in Table 1, AVLA-Collab achieves the best performance in Maze, with a 26.16% success rate. This result is mainly consistent with the observation in Ranch: navigation in complex scenes requires complementary information from multiple modalities. CRONA performs only moderately worse than the homogeneous collaboration baselines, suggesting that cross-modal collaboration can still exploit partial, modality-specific inputs effectively. This finding, together with its best overall performance (Figure 3(f)), indicates that CRONA provides a robust and efficient alternative to multi-modal collaboration. The success rates for all scenes are shown in Appendix D. Table 1: Comparison between CRONA and baselines across five scenes: Studio, Apartment, Ranch, Corridor, and Maze. Dist, Detect, and Succ denote the average distance from each agent to its nearest target object (m), target detection rate (%), and task success rate (%), respectively. Steps and Timeout denote the average number of steps used per episode and the episode timeout rate (%), respectively. Underlined bolds denote the best performance across baselines on each domain. (a) Task performance comparison. Method Studio Corridor Apartment Ranch Maze Dist Succ Dist Succ Dist Detect Succ Dist Detect Succ Dist Detect Succ Single-Agent 3.24 32.66 11.95 5.71 8.58 0.84 31.55 8.68 0.74 12.34 7.29 0.18 0.00 VLA-Collab 1.49 93.65 9.28 14.54 2.32 1.78 78.96 5.75 0.89 38.97 6.89 1.06 18.96 ALA-Collab 3.05 88.17 8.64 25.31 4.34 1.47 38.23 5.33 1.28 42.15 6.81 1.17 19.63 AVLA-Collab 2.91 85.87 9.75 14.29 3.93 1.61 63.38 6.87 0.78 18.93 6.77 1.46 26.16 black!5 CRONA 1.45 95.72 9.11 21.50 3.64 1.69 68.52 5.02 1.58 64.62 7.06 0.93 12.13 (b) Navigation efficiency comparison. Method Studio Corridor Apartment Ranch Maze Steps Timeout Steps Timeout Steps Timeout Steps Timeout Steps Timeout Single-Agent 23.40 1.38 146.58 95.86 434.60 56.47 260.11 18.94 15.32 0.00 VLA-Collab 19.47 0.71 118.92 87.94 289.76 16.08 318.67 22.41 129.13 0.79 ALA-Collab 20.18 0.85 95.66 74.68 342.59 40.32 490.26 32.87 156.18 0.88 AVLA-Collab 21.59 0.92 116.34 87.36 308.27 28.96 396.28 24.88 624.50 20.45 black!5 CRONA 16.08 0.65 135.51 88.55 293.51 24.87 606.53 36.90 293.69 7.14 5.4 Ablation Study Table 2 analyzes the effects of model capacity, input-signal quality, and framework components. We vary the embedding size and compare the homogeneous multi-modal baseline AVLA-Collab with CRONA. AVLA-Collab is highly sensitive to representation capacity. With a small embedding size, agents must compress visual, acoustic, and language information into a limited latent space, leading to poor collaboration performance (0.06% success at embedding size 60). Increasing the embedding size adds only modest overhead (roughly 1 MiB for every additional 40 dimensions), but improves success rate by up to 29.61%. With sufficient capacity, AVLA-Collab can even outperform CRONA at the same embedding size. This suggests that full-modality agents can benefit from rich inputs once capacity is no longer the bottleneck. CRONA is more stable across embedding sizes, since each agent processes fewer modalities and faces a simpler representation-learning problem. We vary image resolution to evaluate robustness to visual signal quality. The homogeneous vision-based methods, VLA-Collab and AVLA-Collab, degrade substantially at low resolution, achieving only 12.76% and 15.43% success, respectively. CRONA is more robust, maintaining 42.76%-65.48% success across different resolutions. This robustness comes from modality specialization. Even with poor visual observations, the audio-based agent may take over and help to maintain the performance. Finally, we ablate the auxiliary beliefs and state input of the centralized critic in Table 2c. Removing the category belief only slightly reduces performance, by 0.72% for AVLA-Collab and 2.04% for CRONA. In contrast, the location belief has a much larger effect. Once it is removed, either alone or together with the category belief, the success rate drops by about half for both methods. We find state information crucial for centralized training, where both methods almost fail (less than 0.2% success rate) to learn without it. Overall, each component contributes to performance, with location belief and state information playing the most important roles. Table 2: Ablation studies on Ranch. (a) compares AVLA-Collab and CRONA across embedding sizes, reporting model size (MiB) and task success rate (%). (b) compares VLA-Collab, AVLA-Collab, and CRONA across visual resolutions, reporting task success rate (%) and steps per episode. (c) compares AVLA-Collab with CRONA and ablates key CRONA components. † denotes the pivot setting used in Table 1. Subscripted arrows show absolute changes relative to the corresponding † pivot entry, where ↑ denotes an increase and ↓ denotes a decrease. Underlined bolds mark the best performance under each setting. (a) Embedding-size ablation. Method 60 100† 140 180 Size Succ Size Succ Size Succ Size Succ AVLA-Collab 36.95↓0.94_\; [rgb]0.15234375,0.4296875,0.6875 \,0.94 0.06↓18.87_\; [rgb]0.15234375,0.4296875,0.6875 \,18.87 37.89† 18.93† 38.83↑0.94_\; [rgb]0.71875,0.234375,0.2265625 \,0.94 43.72↑24.79_\; [rgb]0.71875,0.234375,0.2265625 \,24.79 39.76↑1.87_\; [rgb]0.71875,0.234375,0.2265625 \,1.87 73.33↑54.40_\; [rgb]0.71875,0.234375,0.2265625 \,54.40 black!5 CRONA 27.11↓0.93_\; [rgb]0.15234375,0.4296875,0.6875 \,0.93 11.38↓53.24_\; [rgb]0.15234375,0.4296875,0.6875 \,53.24 28.04† 64.62† 28.98↑0.94_\; [rgb]0.71875,0.234375,0.2265625 \,0.94 65.54↑0.92_\; [rgb]0.71875,0.234375,0.2265625 \,0.92 29.92↑1.88_\; [rgb]0.71875,0.234375,0.2265625 \,1.88 68.75↑4.13_\; [rgb]0.71875,0.234375,0.2265625 \,4.13 (b) Resolution ablation. Method 4 × 4 8 × 8 16 × 16† 32 × 32 Succ Steps Succ Steps Succ Steps Succ Steps VLA-Collab 12.76↓26.21_\; [rgb]0.15234375,0.4296875,0.6875 \,26.21 317.68↓0.99_\; [rgb]0.15234375,0.4296875,0.6875 \,0.99 16.51↓22.46_\; [rgb]0.15234375,0.4296875,0.6875 \,22.46 343.08↑24.41_\; [rgb]0.71875,0.234375,0.2265625 \,24.41 38.97† 318.67† 63.53↑24.56_\; [rgb]0.71875,0.234375,0.2265625 \,24.56 581.70↑263.03_\; [rgb]0.71875,0.234375,0.2265625 \,263.03 AVLA-Collab 15.43↓3.50_\; [rgb]0.15234375,0.4296875,0.6875 \,3.50 320.76↓75.52_\; [rgb]0.15234375,0.4296875,0.6875 \,75.52 18.25↓0.68_\; [rgb]0.15234375,0.4296875,0.6875 \,0.68 322.65↓73.63_\; [rgb]0.15234375,0.4296875,0.6875 \,73.63 18.93† 396.28† 19.21↑0.28_\; [rgb]0.71875,0.234375,0.2265625 \,0.28 388.29↓7.99_\; [rgb]0.15234375,0.4296875,0.6875 \,7.99 black!5 CRONA 42.76↓21.86_\; [rgb]0.15234375,0.4296875,0.6875 \,21.86 346.19↓260.34_\; [rgb]0.15234375,0.4296875,0.6875 \,260.34 62.04↓2.58_\; [rgb]0.15234375,0.4296875,0.6875 \,2.58 573.81↓32.72_\; [rgb]0.15234375,0.4296875,0.6875 \,32.72 64.62† 606.53† 65.48↑0.86_\; [rgb]0.71875,0.234375,0.2265625 \,0.86 615.92↑9.39_\; [rgb]0.71875,0.234375,0.2265625 \,9.39 (c) Component ablation. Method w/o Category Belief w/o Location Belief w/o Any Belief Critic w/o State Full† AVLA-Collab 18.21↓0.72_\; [rgb]0.15234375,0.4296875,0.6875 \,0.72 8.78↓10.15_\; [rgb]0.15234375,0.4296875,0.6875 \,10.15 8.75↓10.18_\; [rgb]0.15234375,0.4296875,0.6875 \,10.18 0.06↓18.87_\; [rgb]0.15234375,0.4296875,0.6875 \,18.87 18.93† black!5 CRONA 62.58↓2.04_\; [rgb]0.15234375,0.4296875,0.6875 \,2.04 26.16↓38.46_\; [rgb]0.15234375,0.4296875,0.6875 \,38.46 31.40↓33.22_\; [rgb]0.15234375,0.4296875,0.6875 \,33.22 0.13↓64.49_\; [rgb]0.15234375,0.4296875,0.6875 \,64.49 64.62† 6 Conclusion We propose CRONA, a decentralized MARL framework for cross-modal navigation. By assigning complementary sensory modalities to different agents, CRONA reduces the burden of learning dense multi-modal representations within each agent, while retaining fully decentralized execution at test time. Experiments show that homogeneous collaboration with limited modalities may suffice for short-range navigation, while heterogeneous collaboration with complementary modalities generally performs better. In more complex scenes, richer multi-modal inputs and sufficient model capacity are also important for navigation. Overall, cross-modal collaboration is a robust and efficient alternative to multi-modal collaboration, especially when targets exhibit clean, modality-specific cues. Limitations This work has several limitations that suggest directions for future exploration. First, we focus on two common modalities, vision and audio, and extending CRONA to other sensory inputs, such as point clouds, LiDAR, or tactile signals, requires further study. Second, we use location and category beliefs as a proof of concept for auxiliary belief learning. Developing control-relevant belief representations for broader modalities and task structures is an important direction. Finally, due to constraints in the environment configuration, our current tasks are instantiated in 2D navigation settings. Extending cross-modal collaboration to full 3D embodied environments would further test its generality and practical applicability. References [1] S. V. Albrecht, F. Christianos, and L. Schäfer (2024) Multi-agent reinforcement learning: foundations and modern approaches. MIT Press. External Links: Link Cited by: §2, §4.3. [2] C. Amato (2024) An introduction to centralized training for decentralized execution in cooperative multi-agent reinforcement learning. arXiv preprint arXiv:2409.03052. Cited by: §2, §4.3. [3] P. Anderson, Q. Wu, D. Teney, J. Bruce, M. Johnson, N. Sünderhauf, I. Reid, S. Gould, and A. Van Den Hengel (2018) Vision-and-language navigation: interpreting visually-grounded navigation instructions in real environments. In Proceedings of the IEEE conference on computer vision and pattern recognition, p. 3674–3683. Cited by: §1, §2. [4] R. Azzam, I. Boiko, and Y. Zweiri (2023) Swarm cooperative navigation using centralized training and decentralized execution. Drones 7 (3), p. 193. Cited by: §1, §2. [5] A. Brohan, N. Brown, J. Carbajal, Y. Chebotar, J. Dabis, C. Finn, K. Gopalakrishnan, K. Hausman, A. Herzog, J. Hsu, et al. (2022) Rt-1: robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817. Cited by: §1. [6] J. R. Bruce and M. M. Veloso (2006) Safe multirobot navigation within dynamics constraints. Proceedings of the IEEE 94 (7), p. 1398–1411. Cited by: §2. [7] W. Burgard, M. Moors, C. Stachniss, and F. E. Schneider (2005) Coordinated multi-robot exploration. IEEE Transactions on robotics 21 (3), p. 376–386. Cited by: §1. [8] A. Chang, A. Dai, T. Funkhouser, M. Halber, M. Niessner, M. Savva, S. Song, A. Zeng, and Y. Zhang (2017) Matterport3D: learning from rgb-d data in indoor environments. International Conference on 3D Vision (3DV). Cited by: §5. [9] C. Chen, Z. Al-Halah, and K. Grauman (2021) Semantic audio-visual navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 15516–15525. Cited by: §2. [10] C. Chen, U. Jain, C. Schissler, S. V. A. Gari, Z. Al-Halah, V. K. Ithapu, P. Robinson, and K. Grauman (2020) Soundspaces: audio-visual navigation in 3d environments. In European conference on computer vision, p. 17–36. Cited by: §B.2, §B.3, §1, §2, §5.2. [11] C. Chen, S. Majumder, Z. Al-Halah, R. Gao, S. K. Ramakrishnan, and K. Grauman (2020) Learning to set waypoints for audio-visual navigation. arXiv preprint arXiv:2008.09622. Cited by: §2. [12] S. Chen, P. Guhur, C. Schmid, and I. Laptev (2021) History aware multimodal transformer for vision-and-language navigation. Advances in neural information processing systems 34, p. 5834–5847. Cited by: §2. [13] X. Chen, H. Ma, J. Wan, B. Li, and T. Xia (2017) Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, p. 1907–1915. Cited by: §1. [14] Y. Cheng and G. Y. Wang (2018) Mobile robot navigation based on lidar. In 2018 Chinese control and decision conference (CCDC), p. 1243–1246. Cited by: §2. [15] C. Claus and C. Boutilier (1998) The dynamics of reinforcement learning in cooperative multiagent systems. AAAI/IAAI 1998 (746-752), p. 2. Cited by: §2. [16] A. Das, S. Kottur, K. Gupta, A. Singh, D. Yadav, J. M. Moura, D. Parikh, and D. Batra (2017) Visual dialog. In Proceedings of the IEEE conference on computer vision and pattern recognition, p. 326–335. Cited by: §1. [17] D. Driess, F. Xia, M. S. Sajjadi, C. Lynch, A. Chowdhery, B. Ichter, A. Wahid, J. Tompson, Q. Vuong, T. Yu, et al. (2023) Palm-e: an embodied multimodal language model. arXiv preprint arXiv:2303.03378. Cited by: §1. [18] J. Duan, S. Yu, H. L. Tan, H. Zhu, and C. Tan (2022) A survey of embodied ai: from simulators to research tasks. IEEE Transactions on Emerging Topics in Computational Intelligence 6 (2), p. 230–244. Cited by: §1. [19] J. Foerster, G. Farquhar, T. Afouras, N. Nardelli, and S. Whiteson (2018-04) Counterfactual multi-agent policy gradients. In Proceedings of the AAAI Conference on Artificial Intelligence, Cited by: §2. [20] D. Fried, R. Hu, V. Cirik, A. Rohrbach, J. Andreas, L. Morency, T. Berg-Kirkpatrick, K. Saenko, D. Klein, and T. Darrell (2018) Speaker-follower models for vision-and-language navigation. In Advances in neural information processing systems, Vol. 31. Cited by: §2. [21] S. Gu, M. Geng, and L. Lan (2021) Attention-based fault-tolerant approach for multi-agent reinforcement learning systems. Entropy 23 (9), p. 1133. Cited by: §1. [22] H. Hao, M. Han, C. Li, Z. Li, and X. Chang (2025) Conav: collaborative cross-modal reasoning for embodied navigation. arXiv preprint arXiv:2505.16663. Cited by: §1, §2, §5.2. [23] W. Hao, C. Li, X. Li, L. Carin, and J. Gao (2020) Towards learning a generic agent for vision-and-language navigation via pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 13137–13146. Cited by: §2. [24] K. He, X. Zhang, S. Ren, and J. Sun (2016) Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, p. 770–778. Cited by: §C.2, §4.2. [25] Y. Hong, Q. Wu, Y. Qi, C. Rodriguez-Opazo, and S. Gould (2021) Vln bert: a recurrent vision-and-language bert for navigation. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, p. 1643–1653. Cited by: §2. [26] Y. Hu, K. Wang, M. Liu, H. Tang, and L. Nie (2023) Semantic collaborative learning for cross-modal moment localization. ACM Transactions on Information Systems 42 (2), p. 1–26. Cited by: §1, §2. [27] C. Huang, O. Mees, A. Zeng, and W. Burgard (2023) Visual language maps for robot navigation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), p. 10608–10615. Cited by: §2. [28] H. Huang, A. V. Savkin, and C. Huang (2021) Decentralized autonomous navigation of a uav network for road traffic monitoring. IEEE Transactions on Aerospace and Electronic Systems 57 (4), p. 2558–2564. Cited by: §1, §2. [29] Y. Huang, J. Lin, C. Zhou, H. Yang, and L. Huang (2022) Modality competition: what makes joint training of multi-modal network fail in deep learning?(provably). In International conference on machine learning, p. 9226–9259. Cited by: §1, §2. [30] S. Iqbal and F. Sha (2019) Actor-attention-critic for multi-agent reinforcement learning. In International conference on machine learning, p. 2961–2970. Cited by: §2. [31] R. Janssen, R. van de Molengraft, H. Bruyninckx, and M. Steinbuch (2016) Cloud based centralized task control for human domain multi-robot operations. Intelligent Service Robotics 9 (1), p. 63–77. Cited by: §2. [32] C. Jia, Y. Yang, Y. Xia, Y. Chen, Z. Parekh, H. Pham, Q. Le, Y. Sung, Z. Li, and T. Duerig (2021) Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, p. 4904–4916. Cited by: §1. [33] W. Kim, B. Son, and I. Kim (2021) Vilt: vision-and-language transformer without convolution or region supervision. In International conference on machine learning, p. 5583–5594. Cited by: §1. [34] E. Kolve, R. Mottaghi, W. Han, E. VanderBilt, L. Weihs, A. Herrasti, M. Deitke, K. Ehsani, D. Gordon, Y. Zhu, et al. (2017) Ai2-thor: an interactive 3d environment for visual ai. arXiv preprint arXiv:1712.05474. Cited by: §1. [35] J. Krantz, E. Wijmans, A. Majumdar, D. Batra, and S. Lee (2020) Beyond the nav-graph: vision-and-language navigation in continuous environments. In European Conference on Computer Vision, p. 104–120. Cited by: §2. [36] J. Ku, M. Mozifian, J. Lee, A. Harakeh, and S. L. Waslander (2018) Joint 3d proposal generation and object detection from view aggregation. In 2018 IEEE/RSJ international conference on intelligent robots and systems (IROS), p. 1–8. Cited by: §1. [37] J. Li, D. Li, S. Savarese, and S. Hoi (2023) Blip-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, p. 19730–19742. Cited by: §1, §2. [38] R. Liu, Y. Shen, P. Gao, P. Tokekar, and M. Lin (2025) Caml: collaborative auxiliary modality learning for multi-agent systems. arXiv preprint arXiv:2502.17821. Cited by: §1, §2, §2, §5.2. [39] R. Lowe, Y. I. Wu, A. Tamar, J. Harb, O. Pieter Abbeel, and I. Mordatch (2017) Multi-agent actor-critic for mixed cooperative-competitive environments. Advances in neural information processing systems 30. Cited by: §2. [40] R. Luna and K. E. Bekris (2011) Efficient and complete centralized multi-robot path planning. In 2011 IEEE/RSJ International Conference on Intelligent Robots and Systems, p. 3268–3275. Cited by: §2. [41] X. Lyu, A. Baisero, Y. Xiao, B. Daley, and C. Amato (2023) On centralized critics in multi-agent reinforcement learning. Journal of Artificial Intelligence Research 77, p. 295–354. Cited by: Appendix A, §2, §4.3. [42] X. Lyu, Y. Xiao, B. Daley, and C. Amato (2021) Contrasting centralized and decentralized critics in multi-agent reinforcement learning. arXiv preprint arXiv:2102.04402. Cited by: §2. [43] A. Majumdar, A. Shrivastava, S. Lee, P. Anderson, D. Parikh, and D. Batra (2020) Improving vision-and-language navigation with image-text pairs from the web. In European Conference on Computer Vision, p. 259–274. Cited by: §2. [44] Manolis Savva*, Abhishek Kadian*, Oleksandr Maksymets*, Y. Zhao, E. Wijmans, B. Jain, J. Straub, J. Liu, V. Koltun, J. Malik, D. Parikh, and D. Batra (2019) Habitat: A Platform for Embodied AI Research. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), Cited by: §1, §5. [45] B. McFee, C. Raffel, D. Liang, D. P. Ellis, M. McVicar, E. Battenberg, O. Nieto, et al. (2015) Librosa: audio and music signal analysis in python.. SciPy 2015 (18-24), p. 7. Cited by: §5. [46] F. A. Oliehoek, M. T. Spaan, and N. Vlassis (2008) Optimal and approximate q-value functions for decentralized pomdps. Journal of Artificial Intelligence Research 32, p. 289–353. Cited by: §1. [47] F. A. Oliehoek and C. Amato (2016) A concise introduction to decentralized pomdps. Springer. External Links: Document Cited by: §1, §3.1. [48] L. E. Parker (2002) ALLIANCE: an architecture for fault tolerant multirobot cooperation. IEEE transactions on robotics and automation 14 (2), p. 220–240. Cited by: §1. [49] S. Paul, A. Roy-Chowdhury, and A. Cherian (2022) Avlen: audio-visual-language embodied navigation in 3d environments. Advances in Neural Information Processing Systems 35, p. 6236–6249. Cited by: §2. [50] X. Peng, Y. Wei, A. Deng, D. Wang, and D. Hu (2022) Balanced multimodal learning via on-the-fly gradient modulation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 8238–8247. Cited by: §1, §2. [51] L. Peshkin, K. Kim, N. Meuleau, and L. P. Kaelbling (2001) Learning to cooperate via policy search. arXiv preprint cs/0105032. Cited by: §2. [52] C. R. Qi, W. Liu, C. Wu, H. Su, and L. J. Guibas (2018) Frustum pointnets for 3d object detection from rgb-d data. In Proceedings of the IEEE conference on computer vision and pattern recognition, p. 918–927. Cited by: §1. [53] Y. Qi, Z. Pan, Y. Hong, M. Yang, A. Van Den Hengel, and Q. Wu (2021) The road to know-where: an object-and-room informed sequential bert for indoor vision-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 1655–1664. Cited by: §2. [54] T. Qin, M. Macdonald, and D. Qiao (2021) Fully decentralized cooperative navigation for spacecraft constellations. IEEE Transactions on Aerospace and Electronic Systems 57 (4), p. 2383–2394. Cited by: §1, §2. [55] M. Shridhar, L. Manuelli, and D. Fox (2022) Cliport: what and where pathways for robotic manipulation. In Conference on robot learning, p. 894–906. Cited by: §1. [56] M. Shridhar, J. Thomason, D. Gordon, Y. Bisk, W. Han, R. Mottaghi, L. Zettlemoyer, and D. Fox (2020) Alfred: a benchmark for interpreting grounded instructions for everyday tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 10740–10749. Cited by: §1. [57] M. Shridhar, X. Yuan, M. Côté, Y. Bisk, A. Trischler, and M. Hausknecht (2020) Alfworld: aligning text and embodied environments for interactive learning. arXiv preprint arXiv:2010.03768. Cited by: §1. [58] R. Simmons, D. Apfelbaum, W. Burgard, D. Fox, M. Moors, S. Thrun, and H. Younes (2000) Coordination for multi-robot exploration and mapping. In Aaai/Iaai, p. 852–858. Cited by: §1. [59] A. Szot, A. Clegg, E. Undersander, E. Wijmans, Y. Zhao, J. Turner, N. Maestre, M. Mukadam, D. Chaplot, O. Maksymets, A. Gokaslan, V. Vondrus, S. Dharur, F. Meier, W. Galuba, A. Chang, Z. Kira, V. Koltun, J. Malik, M. Savva, and D. Batra (2021) Habitat 2.0: training home assistants to rearrange their habitat. In Advances in Neural Information Processing Systems (NeurIPS), Cited by: §1, §5. [60] M. Tan et al. (1993) Multi-agent reinforcement learning: independent vs. cooperative agents. In Proceedings of the tenth international conference on machine learning, p. 330–337. Cited by: §2. [61] Y. H. Tsai, S. Bai, P. P. Liang, J. Z. Kolter, L. Morency, and R. Salakhutdinov (2019) Multimodal transformer for unaligned multimodal language sequences. In Proceedings of the 57th annual meeting of the association for computational linguistics, p. 6558–6569. Cited by: §1. [62] K. Tuyls, K. Verbeeck, and T. Lenaerts (2003) A selection-mutation model for q-learning in multi-agent systems. In Proceedings of the second international joint conference on Autonomous agents and multiagent systems, p. 693–700. Cited by: §2. [63] J. van Den Berg, J. Snoeyink, M. C. Lin, and D. Manocha (2009) Centralized path planning for multiple robots: optimal decoupling into sequential plans.. In Robotics: Science and systems, Vol. 2, p. 2–3. Cited by: §2. [64] P. Velagapudi, K. Sycara, and P. Scerri (2010) Decentralized prioritized planning in large multirobot teams. In 2010 IEEE/RSJ International Conference on Intelligent Robots and Systems, p. 4603–4609. Cited by: §2. [65] H. Wang, W. Wang, X. Zhu, J. Dai, and L. Wang (2021) Collaborative visual navigation. arXiv preprint arXiv:2107.01151. Cited by: §1. [66] T. Wang, X. Li, F. Lu, T. Gong, J. Dong, W. Xue, S. Qu, C. Bai, and G. Chen (2026) CoNavBench: collaborative long-horizon vision-language navigation benchmark. In The Fourteenth International Conference on Learning Representations, Cited by: §1, §2, §5.2. [67] W. Wang, D. Tran, and M. Feiszli (2020) What makes training multi-modal classification networks hard?. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 12695–12705. Cited by: §1, §2. [68] W. Wang, L. Mao, R. Wang, and B. Min (2024) Multi-robot cooperative socially-aware navigation using multi-agent reinforcement learning. In 2024 IEEE International Conference on Robotics and Automation (ICRA), p. 12353–12360. Cited by: §1, §2. [69] N. Wu, S. Jastrzebski, K. Cho, and K. J. Geras (2022) Characterizing and overcoming the greedy nature of learning in multi-modal deep neural networks. In International Conference on Machine Learning, p. 24043–24055. Cited by: §1, §2. [70] M. Wunder, M. L. Littman, and M. Babes (2010) Classes of multiagent q-learning dynamics with epsilon-greedy exploration. In Proceedings of the 27th International Conference on Machine Learning (ICML-10), p. 1167–1174. Cited by: §2. [71] Y. Xiao, J. Hoffman, T. Xia, and C. Amato (2020) Learning multi-robot decentralized macro-action-based policies via a centralized q-net. In 2020 IEEE International conference on robotics and automation (ICRA), p. 10695–10701. Cited by: §1, §2. [72] Y. Xiao, W. Tan, and C. Amato (2022) Asynchronous actor-critic for multi-agent reinforcement learning. Advances in Neural Information Processing Systems 35, p. 4385–4400. Cited by: §1, §2. [73] Y. Xue and W. Chen (2023) Multi-agent deep reinforcement learning for uavs navigation in unknown complex environment. IEEE Transactions on Intelligent Vehicles 9 (1), p. 2290–2303. Cited by: §1, §2. [74] S. Yao, R. Guan, X. Huang, Z. Li, X. Sha, Y. Yue, E. G. Lim, H. Seo, K. L. Man, X. Zhu, et al. (2023) Radar-camera fusion for object detection and semantic segmentation in autonomous driving: a comprehensive review. IEEE Transactions on Intelligent Vehicles 9 (1), p. 2094–2128. Cited by: §2. [75] Y. Yao, T. Yu, A. Zhang, C. Wang, J. Cui, H. Zhu, T. Cai, H. Li, W. Zhao, Z. He, et al. (2024) Minicpm-v: a gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800. Cited by: §1. [76] B. Yu, H. Kasaei, and M. Cao (2023) L3mvn: leveraging large language models for visual target navigation. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), p. 3554–3560. Cited by: §2. [77] C. Yu, A. Velu, E. Vinitsky, J. Gao, Y. Wang, A. Bayen, and Y. Wu (2022) The surprising effectiveness of ppo in cooperative multi-agent games. In Advances in Neural Information Processing Systems, Vol. 35, p. 24611–24624. Cited by: §2. [78] L. Yuan, Z. Zhang, L. Li, C. Guan, and Y. Yu (2023) A survey of progress on cooperative multi-agent reinforcement learning in open environment. arXiv preprint arXiv:2312.01058. Cited by: §2. [79] G. Yuksel, M. van Gerven, and K. van der Heijden (2026) GRAM: spatial general-purpose audio representations for real-world environments. arXiv preprint arXiv:2602.03307. Cited by: §2. [80] H. Zhang, Y. Yu, L. Wang, F. Sun, and W. Zheng (2025) Advancing audio-visual navigation through multi-agent collaboration in 3d environments. In International Conference on Neural Information Processing, p. 502–516. Cited by: §1, §2, §5.2. [81] K. Zhang, Z. Yang, and T. Başar (2021) Multi-agent reinforcement learning: a selective overview of theories and algorithms. Handbook of reinforcement learning and control, p. 321–384. Cited by: §2. Appendix A Proofs Proposition 1. Assume that the belief predictor is accurate, such that the predicted belief \,b is a correct history-induced latent representation of \,h. Then the state-history-belief value V(,,s)\,V π(h,b,s) provides an unbiased estimate of V()\,V π(h). Proof. By Lemma 2 of [41], the state-augmented action-value function is unbiased with respect to the history-conditioned value, i.e., Q(,)=s∣[Q(,s,)].Q π(h,a)=E_s [Q π(h,s,a) ]. Since b is inferred from h and is assumed to be correct, conditioning on (,)(h,b) does not introduce additional information beyond the history. Therefore, p(s∣,)=p(s∣),p(s ,b)=p(s ), and the augmented critic satisfies Q(,,s,)=Q(,s,).Q π(h,b,s,a)=Q π(h,s,a). Then, V() V π(h) =∼(⋅∣)[Q(,)] =E_a π(· ) [Q π(h,a) ] =∼(⋅∣)[s∣[Q(,s,)]] =E_a π(· ) [E_s [Q π(h,s,a) ] ] =s∣[∼(⋅∣)[Q(,,s,)]] =E_s [E_a π(· ) [Q π(h,b,s,a) ] ] =s∣[V(,,s)]. =E_s [V π(h,b,s) ]. Thus, when s is sampled from the posterior state distribution p(s∣)p(s ), the estimator V(,,s)V π(h,b,s) has expectation V()V π(h). Therefore, V(,,s)V π(h,b,s) is an unbiased estimator of V()V π(h). ∎ Appendix B Collaborative Navigation Benchmark B.1 Scenes The bird’s-eye-view of the scenes we used are shown in Figure 4. Table 3 shows the Matterport3D scene IDs, the number of navigable points, and the total navigable area of each scene. (a) Studio ∣ Picture (b) Corridor ∣ Sink (c) Apartment ∣ Bed & Counter (d) Ranch ∣ Picture & Table (e) Maze ∣ Drawer & Table & Chair Figure 4: Bird-eye’s-views of MatterPort3D scenes. B.2 Dataset Construction Details To construct datasets, we use base episodes in [10] according to the desired Matterport3D scenes. We chose 1, 2, or 3 targets from different categories from the initial datasets. Episodes are filtered out if the initial distances between agents, or between agents and target objects are below or exceed predefined thresholds. Table 4 summarizes the dataset construction parameters. Table 3: Statistics of the five scenes used in our collaborative navigation dataset. Scene ID in Matterport3D Navigable points Navigable area (m2) Studio GdvgFV5R1Z5 20 20.49 Corridor ac26ZMwG7aT 619 369.33 Apartment 17DRP5sb8fy 83 52.04 Ranch JeFG25nYj2p 193 166.22 Maze B6ByNegPMKs 1603 1348.31 Some Matterport3D scenes contain disconnected navigable regions. To create diverse but valid episodes, we filter the generated tasks using both minimum- and maximum-distance constraints. The minimum-distance constraint avoids overly easy episodes where agents start too close to the targets, while the maximum-distance constraint removes episodes in which some targets are unreachable. We apply these constraints in Corridor and Maze, where disconnected regions occur more frequently. As a result, although Corridor contains a high number of navigable areas, a horizon of 150 is enough for all Corridor episodes. Table 4: Dataset construction parameters for the five chosen MatterPort3D scenes. A single-object dataset keeps one target per episode, while the multi-object dataset combines multiple targets. Scene Targets Target dist. Start-goal dist. Train eps. Studio picture - ≥2.0m≥ 2.0\,m 220 Corridor sink - 22-5m5\,m 218 Apartment bed, counter ≥2.0m≥ 2.0\,m ≥4.0m≥ 4.0\,m 230 Ranch picture, table ≥2.0m≥ 2.0\,m ≥4.0m≥ 4.0\,m 228 Maze chair, table, chest_of_drawers 33-10m10\,m 33-10m10\,m 252 Two agents are initialized at different starting positions. For multi-object datasets, an episode is considered successful if each target is reached within a distance threshold of 1m1\,m. For two-object datasets, agents must start at least 1.5m1.5\,m apart. For Corridor, the initial distance lie between 2.02.0 and 5.0m5.0\,m. For Maze, the initial distance lie between 3.03.0 and 10.0m10.0\,m. We split each dataset into training and validation sets with 3:13:1. B.3 Acoustic Simulation We use shared material configurations across all evaluated scenes to simulate the physical acoustics of each space. Each material contains frequency-dependent absorption, scattering, and transmission coefficients that simulate the sound propagation. The configuration here only affects acoustic rendering and does not change navigable points or the navigable area. Table 5 shows mappings between semantics and materials, with the corresponding coefficient ranges. Table 5: Representative acoustic material configurations for audio rendering. Acoustic material Example semantic labels Absorption Scattering Transmission Acoustic Tile ceiling 0.50-0.70 0.10-0.30 0.002-0.050 Gypsum Board wall 0.04-0.29 0.10-0.15 0.001-0.035 Carpet floor, mat 0.01-0.65 0.10-0.45 0.001-0.008 Glass window, mirror, tv_monitor 0.05-0.35 0.05-0.05 0.022-0.125 Foliage plant, indoor-plant 0.03-0.31 0.20-0.80 0.30-0.90 Steel sink, microwave, railing 0.02-0.10 0.10-0.10 0.056-0.250 wood, Thick chair, table, counter 0.05-0.19 0.10-0.15 0.001-0.035 Wood Floor cabinet, stair 0.06-0.15 0.10-0.15 0.002-0.071 Curtain bed, blanket, cushion, sofa 0.07-0.75 0.10-0.50 0.045-0.420 Default default 0.10-0.10 0.50-0.50 0.000-0.000 We choose the following sounds from [10] for constructing our dataset: • Dragging Chair • Table with Silverware Dropping • Picture with Camera Shutter • Sink with Dripping Water • Counter with Coin Drop • Pulling Chest of Drawers • Creaking Bed Figure 5 shows the corresponding source spectrograms. font=scriptsize (a) Dragging Chair (b) Table with Silverware (c) Picture Shutter (d) Sink Dripping (e) Coin Drop on Counter (f) Chest of Drawers (g) Creaking Bed Figure 5: Spectrograms of selected sounds above. Agents learn more effectively from short, clean, well-isolated sounds. Such sounds usually exhibit a sharp attack, little or no sustain, and a rapid decay with minimal trailing energy. For example, Sink and Table clearly follow this structure, producing consistent acoustic patterns that provide reliable cues for target localization. (a) Studio ∣ Picture (b) Corridor ∣ Sink (c) Apartment ∣ Bed & Counter (d) Ranch ∣ Picture & Table (e) Maze ∣ Drawer & Table & Chair Figure 6: Illustration of example episodes. B.4 Episodes For each episode, objects will be selected based on the corresponding categories. Each episode strictly follows Table 4, the demonstrations of example episodes are in Figure 6. Appendix C Experimental Settings C.1 Hyperparameters We use the same hyperparameters across all scenes. Table 6 summarizes the training configurations. Table 6: Shared training hyperparameters used across all evaluated scenes. Hyperparameter Value Hyperparameter Value Optimizer Adam Actor learning rate 0.00025 Critic learning rate 0.0002 Adam epsilon 1e-5 PPO epochs 2 PPO mini-batches 1 Rollout steps per update 150 PPO policy ratio ρ clip 0.2 PPO value clip 0.25 Policy-value loss coefficient μ 0.67 Belief smooth coefficient α 0.5 Entropy coefficient β 0.05 Discount factor γ 0.99 GAE λ 0.95 Max gradient norm 0.2 History cache size (steps) 150 Vision encoder hidden size 128 Audio encoder hidden size 128 Transformer hidden size 256 Language embedding size 384 Language encoder hidden size 24 Normalized advantage False Number of updates 8000 Max depth for depth sensor 3 m C.2 Architecture Agent Configuration Since many objects in Matterport3D scenes are large and visually salient, a single agent can effortlessly localize them without requiring collaboration. To increase the difficulty of tasks, we restrict visual observations to depth maps with a sensing range of 0–5m5\,m, a resolution of 16×1616× 16 pixels, and a horizontal field of view (HFoV) of 10∘10 . We use sentence-transformers/all-MiniLM-L6-v2 as the instruction encoder, a compact ResNet-18 as the visual encoder [24], and a plain CNN with 3 convolutional layers as the audio encoder. At each time step, each agent combines its previous observations and actions with the current observation using an MLP, then encodes them using 88-head transformers to obtain a history representation. Audio Encoder • Input: Binaural spectrogram. • Layers: – Conv2d(2,32,5×5,stride=2)+ReLUConv2d(2,32,5× 5,stride=2)+ReLU – Conv2d(32,64,3×3,stride=2)+ReLUConv2d(32,64,3× 3,stride=2)+ReLU – Conv2d(64,64,3×3,stride=1)Conv2d(64,64,3× 3,stride=1) – Flatten Vision Encoder • Input: Single-channel depth observation. • Layers: – ResizeCenterCrop(64×64)ResizeCenterCrop(64× 64) – Conv2d(1,16,7×7,stride=1,padding=3)+GroupNorm+ReLUConv2d(1,16,7× 7,stride=1,padding=3)+GroupNorm+ReLU – 2×ResidualBlock(16→16,stride=1)2×ResidualBlock(16→ 16,stride=1) – 2×ResidualBlock(16→32,stride=2)2×ResidualBlock(16→ 32,stride=2) – 2×ResidualBlock(32→64,stride=2)2×ResidualBlock(32→ 64,stride=2) – 2×ResidualBlock(64→128,stride=2)2×ResidualBlock(64→ 128,stride=2) – Flatten – Linear(128×8×8,64)Linear(128× 8× 8,64) Auxiliary Belief Predictor • Input: Binaural spectrogram. • Layers: – Conv2d(2,16,7×7,stride=1,padding=3)+GroupNorm+ReLUConv2d(2,16,7× 7,stride=1,padding=3)+GroupNorm+ReLU – 2×ResidualBlock(16→16,stride=1)2×ResidualBlock(16→ 16,stride=1) – 2×ResidualBlock(16→32,stride=2)2×ResidualBlock(16→ 32,stride=2) – 2×ResidualBlock(32→64,stride=2)2×ResidualBlock(32→ 64,stride=2) – 2×ResidualBlock(64→128,stride=2)2×ResidualBlock(64→ 128,stride=2) – Flatten – Linear(4608,2)Linear(4608,2) History Encoder • Input: Each agent’s recent observation-action history. • Layers: – Previous action encoding: Linear(||,16)Linear(|A|,16) – Relative pose encoding: Linear(5,16)Linear(5,16) – Feature fusion: Linear(din,dh)+ReLU+Linear(dh,dh)Linear(d_in,d_h)+ReLU+Linear(d_h,d_h) – Transformer encoder: 11 layer with 88 attention heads – Transformer decoder: 11 layer with 88 attention heads – Feed-forward dimension: dhd_h – Activation: ReLU Language Encoder • Input: Tokenized target category or language instruction. • Layers: – WordPiece tokenization with truncation – Token embeddings + position embeddings + segment embeddings – 6×6× Transformer encoder layers – Each layer uses 1212-head self-attention – Feed-forward network: Linear(384,1536)+GELU+Linear(1536,384)Linear(384,1536)+GELU+Linear(1536,384) – Mean pooling over token embeddings using the attention mask – L2L_2 normalization Appendix D Additional Results D.1 Success Rates We provide the success rates for each scene in Figure 7. In most scenes, the success curves align well with the return curves in Figure 3, leading to the same dominance patterns. This is expected because task success contributes the largest portion of the episode return. However, Maze exhibits a noticeable discrepancy: although AVLA-Collab achieves a much higher return than the other methods, its success rate does not improve to the same extent. This is because agents sometimes stop prematurely after reaching easier targets, which helps avoid large penalties from the distance-progress term but prevents them from completing all targets. Overall, the reward still provides useful optimization signals: it guides agents to navigate toward targets, reduce their distance to the goals, and stop within the target vicinity. font=footnotesize (a) Studio ∣ Picture (b) Corridor ∣ Sink (c) Apartment ∣ Bed & Counter (d) Ranch ∣ Picture & Table (e) Maze ∣ Drawer & Table & Chair Figure 7: Additional evaluation of CRONA and collaborative navigation baselines across 5 Matterport3D scenes: (a)-(e) show the success rate of each scene. The x-axis indicates the environment steps. Curves are smoothed by an exponential moving average. Shadows denote 90% bootstrapped CI. Results are averaged over 5 runs. Appendix E Insturction Design We use three categories of prompt templates. For each data entry, the goal is specified in natural language. Audio-based agents sample prompts from the audio-specific templates, vision-based agents sample from the vision-specific templates, and agents with both modalities may also use the general templates. The target object name and its corresponding sound category are substituted into the selected template. The prompt templates used in our benchmark are listed below. ⬇ General Please help me locate the ... and ... Please help me find where the ... and ... are. Your task is to find the ... and ... for me. Show me the locations of the ... and ... Please find the ... and ... in the environment. Search for the ... and report where they are. Navigate to the ... and the ... Find both target objects: the ... and the ... Audio-Based Agent The environment contains sounds from ... and ... Please locate both sound-emitting objects. I cannot find the ... and ... anymore. They sound like ... and ... Can you help me locate them? Can you hear where the ... is? Listen for the sound of ... and use it to find the ... Find the objects that are making ... and ... sounds. Use the audio cues to locate the ... and ... Follow the sounds associated with ... and ... to find the target objects. The ... produces a ... sound. Please locate it using the sound cue. Vision-Based Agent Based on the visual observation, find the ... and ... Look for the ... and ... in the scene. Use visual cues to locate the ... and ... Search the environment for the visible ... and ... Find the ... by observing its shape and appearance. Watch for visual evidence of the ... and ... Inspect the scene and locate the ... and ... Navigate toward the visually observed ... and ... Appendix F Reward Design All agents share a joint team reward. The reward design differs slightly between single-object and multi-object tasks. Single-object tasks For single-object two-agent runs, each agent receives a per-step reward ri=rslack+ridist+risuccr_i=r^slack+r_i^dist+r_i^succ, where rslack=−0.02r^slack=-0.02 is the per-agent time penalty, ridistr_i^dist is the reduction in distance from agent i to the target between consecutive steps, and risucc=20r_i^succ=20 if agent i calls stop within the success distance of the target. The team reward returned to PPO is the sum of the two agent rewards. Hence, the effective per-step slack penalty is −0.04-0.04 for two-agent episodes. No additional stop penalty is applied in the single-object multi-agent environment. An episode is successful if any agent calls stop near the target. Multi-object tasks For multi-object runs, the team reward is defined as rteam=rslack+∑iridist+rstop+rgoalr^team=r^slack+ _ir_i^dist+r^stop+r^goal, where rslack=−0.02r^slack=-0.02 is a team-level time penalty, ∑iridist _ir_i^dist is the sum of distance-progress rewards over agents, and rstop=−0.2⋅nstopr^stop=-0.2· n_stop penalizes agents that call stop. When a new target is found, the goal reward is scaled by task progress, rgoal=NfoundNtotal⋅20r^goal= N_foundN_total· 20, where NfoundN_found is the number of targets found after the current discovery and NtotalN_total is the total number of targets in the episode. So in a two-target task, the first discovered target gives a reward of 1010, and the second discovered target gives a reward of 2020. If both agents call stop before all goals are found, the episode terminates after one such step, with no additional both-stop termination penalty. Maze reward adjustment We adjusted the reward scale for Maze to encourage long-horizon exploration, as it’s the largest and most complex scene with the most targets in our benchmark. Comparably, agents require more time to explore before receiving valuable visual or acoustic evidence, since acoustic cues are relatively weak and ambiguous when agents are far or occluded from the source. To reduce the cost of exploration and provide denser directional guidance toward target regions, we use a slack penalty of −0.002-0.002, a success reward scale of 3.03.0, a distance reward scale of 2.02.0, and a progressive distance reward scale of 1.51.5. Appendix G Compute Resources Experiments were run on a cluster and local workstations. The runtime of each training depends on the scene size, method, model size, and variable hardware. For a run trained to 500k environment steps, Studio took roughly 88-1010 hours on 5090, Apartment and Ranch took roughly 2020-2525 hours on A100, Corridor and Maze took roughly 3030-4848 hours on A100. • Workstation: – GPU: 1× NVIDIA GeForce RTX 5090, 32 GB VRAM – CPU: AMD Ryzen 9 9950X, 16 cores / 32 threads – System memory: 123 GiB • Cloud Cluster: – GPU: 2× NVIDIA A100-SXM4, 160 GB total VRAM – CPU: AMD EPYC 7513, 32 cores – System memory: 354 GB • Software environment: – CUDA: 12.8 – PyTorch: 2.8.0 Appendix H Broader Impacts This work studies how agents with different sensory modalities contribute to collaborative navigation. We build a collaborative navigation benchmark with simulated vision and audio observations in realistic indoor environments. We identify several modality-dominance patterns and analyze when and why each pattern emerges. Our findings suggest that incorporating more modalities does not always lead to better performance; instead, the usefulness of each modality depends on the scene structure, target properties, and sensory reliability. We further propose a MARL framework for cross-modal collaborative navigation. This work opens the door to studying cross-modal collaboration in embodied multi-agent navigation with multi-agent reinforcement learning.