Paper deep dive
CoLA: Multi-VLA Coordination via Learned Adapters
Ahaskar Kashyap, Jiawen Zong, Boyu Ma, Chuhao Zhou, Jingliang Li, Stefano V. Albrecht, Jianfei Yang
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 10/14/2026, 9:05:56 AM
Summary
The paper introduces CoLA (Coordination via Learned Adapters), a framework for multi-agent coordination using frozen Vision-Language-Action (VLA) backbones. CoLA adds lightweight trainable modules: a latent message encoder, a partner message decoder, and a diffusion-based action head. This allows agents to coordinate via explicit communication channels when information is private (e.g., occluded views) but rely on implicit observation when information is shared. Experiments on handover tasks show CoLA significantly outperforms baselines in partial observability settings while matching performance in fully observable ones, using far fewer trainable parameters than fine-tuning approaches.
Entities (11)
Relation Signals (6)
COLA → achievessuccessrate → 76.0%
confidence 95% · On an occluded tray handover... CoLA reaches 76.0% success
Occluded Tray Handover → requires → Explicit Communication
confidence 95% · On an occluded tray handover... CoLA reaches 76.0% success versus 1.3% with messages zeroed throughout training.
COLA → uses → Frozen VLA Backbone
confidence 95% · CoLA keeps each agent's VLA backbone frozen and trains a lightweight message channel and a per-arm diffusion action head on top
Two-arm Handover (Shared Camera) → doesnotrequire → Explicit Communication
confidence 90% · removing messages has no measurable effect with a shared overhead camera (90.6% vs. 88.3% at two arms)
COLA → generalizesto → Different VLA Backbones
confidence 90% · This architecture generalises across different VLA backbones as well. With a frozen π0.5 backbone...
COLA → outperforms → LoRA-fine-tuned baseline
confidence 90% · CoLA matches the overall success rate of a LoRA-fine-tuned baseline (37.3% vs. 34.7%) using a fraction of the trainable parameters.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Vision-Language-Action (VLA) models are typically deployed as isolated single-agent policies. When several such agents share a workspace, they can coordinate implicitly, by observing one another, or explicitly, by exchanging messages, but it remains unclear when explicit communication becomes necessary. We introduce CoLA (Coordination via Learned Adapters), which keeps each agent's VLA backbone frozen and trains a lightweight message channel and a per-arm diffusion action head on top, with fully decentralised execution. Across three tasks that differ in how much each agent can observe, the channel matters only when information is private. On an occluded tray handover, where the box's colour marker is visible to one agent only, CoLA reaches 76.0% success versus 1.3% with messages zeroed throughout training. On two-arm and three-arm handover tasks, removing messages has no measurable effect with a shared overhead camera (90.6% vs. 88.3% at two arms; 63.5% vs. 60.0% at three), but lowers success on the two-arm task's wrist-only variant (74.8% vs. 64.3%). This architecture generalises across different VLA backbones as well. With a frozen $\pi_{0.5}$ backbone on the same occluded tray task, CoLA matches the overall success rate of a LoRA-fine-tuned baseline (37.3% vs. 34.7%) using a fraction of the trainable parameters. When CoLA places the box, it selects the correct tray 71.8% of the time, whereas the independently fine-tuned baseline, whose placing arm never sees the marker, selects it at chance level (34.7% against 33.3%).
Tags
Links
- Source: https://arxiv.org/abs/2610.15337v1
- Canonical: https://arxiv.org/abs/2610.15337v1
Trouble viewing inline? Open PDF directly →
Full Text
71,871 characters extracted from source content.
Expand or collapse full text
CoLA: Multi-VLA Coordination via Learned Adapters Code: https://github.com/Ahaskar04/CoLA.git Ahaskar Kashyap Affiliation: MARS Lab Affiliation: Nanyang Technological University, Singapore Equal contribution Jiawen Zong Affiliation: Autonomous Agents Research Group Affiliation: Nanyang Technological University, Singapore Equal contribution Boyu Ma Affiliation: MARS Lab Affiliation: Nanyang Technological University, Singapore Equal contribution Chuhao Zhou Affiliation: MARS Lab Affiliation: Nanyang Technological University, Singapore Jingliang Li Affiliation: MARS Lab Affiliation: Nanyang Technological University, Singapore Stefano V. Albrecht Affiliation: Autonomous Agents Research Group Affiliation: Nanyang Technological University, Singapore Jianfei Yang Affiliation: MARS Lab Affiliation: Nanyang Technological University, Singapore Corresponding author Abstract Vision-Language-Action (VLA) models are typically deployed as isolated single-agent policies. When several such agents share a workspace, they can coordinate implicitly, by observing one another, or explicitly, by exchanging messages, but it remains unclear when explicit communication becomes necessary. We introduce CoLA (Coordination via Learned Adapters), which keeps each agent’s VLA backbone frozen and trains a lightweight message channel and a per-arm diffusion action head on top, with fully decentralised execution. Across three tasks that differ in how much each agent can observe, the channel matters only when information is private. On an occluded tray handover, where the box’s colour marker is visible to one agent only, CoLA reaches 76.0%76.0\% success versus 1.3%1.3\% with messages zeroed throughout training. On two-arm and three-arm handover tasks, removing messages has no measurable effect with a shared overhead camera (90.6%90.6\% vs. 88.3%88.3\% at two arms; 63.5%63.5\% vs. 60.0%60.0\% at three), but lowers success on the two-arm task’s wrist-only variant (74.8%74.8\% vs. 64.3%64.3\%). This architecture generalises across different VLA backbones as well. With a frozen π0.5 _0.5 backbone on the same occluded tray task, CoLA matches the overall success rate of a LoRA-fine-tuned baseline (37.3% vs. 34.7%) using a fraction of the trainable parameters. When CoLA places the box, it selects the correct tray 71.8%71.8\% of the time, whereas the independently fine-tuned baseline, whose placing arm never sees the marker, selects it at chance level (34.7%34.7\% against 33.3%33.3\%). †correspondence: jianfei.yang@ntu.edu.sg 1 Introduction Robot foundation models trained on large cross-embodiment datasets now generalise well in single-arm manipulation (Brohan et al., 2023b; Brohan et al., 2023a; Octo Model Team et al., 2024; Kim et al., 2024). However, many real-world robotic deployments inherently require several arms or robots to act together, ranging from dual-arm collaborative handovers to multi-robot assembly and warehouse staging (Krnjaic et al., 2024; Yifan et al., 2025). In such domains, task execution cannot rely exclusively on individual dexterity; success relies on mutual perception, intention signalling, and closed-loop behavioural synchrony between autonomous partners. Despite their dexterity in isolation, current VLA models do not coordinate naturally when co-located in a shared workspace (Dey et al., 2026). Each operates under an egocentric assumption: it perceives the workspace through its own sensors and treats partner arms as unmodelled obstacles. Closing this gap faces three practical constraints. Training multi-arm policies end-to-end (Liu et al., 2025b; Zhao et al., 2023; Fu et al., 2024) requires scarce multi-robot data and discards single-agent pretraining. LLM coordinators (Mandi et al., 2024; Liu et al., 2025a) run at 0.50.5–1Hz1\,Hz, which is too slow for reactive manipulation. And deployed robots often run fixed on-device backbones whose weights are costly or impossible to modify (we defer further discussion of related works to section 2). These constraints jointly define the setting we study, including agents whose backbones stay frozen, coordinating without a centralised arbiter, at real-time control rates. In this setting, agents can coordinate only implicitly, by observing one another, or through an explicit channel added alongside their frozen policies. We therefore ask: when does explicit inter-agent communication improve coordination over implicit observation alone, and how does this depend on task coupling and partial observability? To answer this question, we propose Coordination via Learned Adapters (CoLA). CoLA discards each backbone’s native action head entirely and in its place trains three modular components on top of each frozen VLA backbone: a latent message encoder, a partner message decoder, and a coordination-conditioned action chunking head. The encoder compresses the agent’s readout embedding into a low-dimensional message vector. Each agent sends its encoded message to the other agents via the communication channel. Received messages are then decoded and fused with the agent’s self representation, before being mapped to a coordinated action chunk. The encoders, decoders, and action heads are optimised jointly by behavioural cloning on multi-robot demonstrations. We evaluate CoLA in four settings of increasing shared observability, built on three tasks: an occluded tray handover, where the colour marker is visible only to arm A and arm B must place the box in the matching tray; a two-arm handover with each arm’s own wrist camera only; the same two-arm handover with a shared overhead camera; and a three-arm handover chain, which also shares the overhead camera. All settings use a frozen Octo-Base backbone, and the tray task is additionally tested with a frozen π0.5 _0.5 backbone. Against the strongest decentralised baseline in each setting, head-only or fully fine-tuned, CoLA achieves between 5.55.5 and 53.353.3 percentage points higher success rate, with the largest gain on the occluded tray handover (76.0%76.0\% vs. 22.7%22.7\%). To isolate the channel itself, we retrain CoLA with every message fixed to the zero vector during and after training. Removing the channel collapses success on the occluded tray task (76.0%76.0\% to 1.3%1.3\%) and lowers it under wrist-only observation (74.8%74.8\% to 64.3%64.3\%), but has no measurable effect once the arms share an overhead view (90.6%90.6\% vs. 88.3%88.3\% at two arms and 63.5%63.5\% vs. 60.0%60.0\% at three, both within seed-to-seed variation). Whether a team of frozen VLAs needs an explicit channel is therefore determined by the task’s information structure: the channel is essential when task-relevant state is private to one agent, and redundant when the agents can already observe it. Because the channel is a small module trained on cached features of a frozen backbone, it can be added to already-deployed VLA policies without retraining them, and used only where the information structure calls for it. This makes the choice of when and what to communicate a design question in its own right for multi-robot VLA systems. 2 Related Work Multi-Agent Reinforcement Learning and Emergent Communication. Learned inter-agent communication is well-studied in cooperative MARL (Albrecht et al., 2024). Architectures like CommNet (Sukhbaatar et al., 2016), DIAL (Foerster et al., 2016), and TarMAC (Das et al., 2019) optimise continuous differentiable communication channels jointly with agent policies. Emergent communication shows that symbolic or continuous communication protocols can spontaneously ground coordination cues without direct linguistic supervision (Lazaridou et al., 2017; Havrylov and Titov, 2017; Baker et al., 2020). Alternatively, communication can also be strictly bounded to the action space. Recent approaches, such as Multiagent Control via Action Suggestions (MCAS), demonstrate that sharing suggested joint actions, instead of full observations, allows agents to effectively infer belief states under decentralised partial observability (Asmar and Kochenderfer, 2024). However, these methods train all agent policies from scratch. In CoLA, by contrast, each agent’s feature extractor is a pretrained VLA that stays frozen while the channel is learned. Multi-Robot and Bimanual Manipulation. Prior multi-robot manipulation architectures typically rely on unified centralised controllers. Centralised imitation models such as ACT (Zhao et al., 2023), Mobile ALOHA (Fu et al., 2024), InterACT (Lee et al., 2024), and RDT-1B (Liu et al., 2025b) model multiple limbs within a single concatenated action space. TwinVLA (Im et al., 2026) coordinates twin single-arm models via dense cross-attention, while AnyBimanual (Lu et al., 2025) transfers unimanual policies to bimanual control; both, however, update backbone weights. Closer to our setting, CHORUS (Doshi et al., 2026) adapts a single VLA backbone so that each robot runs an independent copy conditioned only on its own observations and a robot-identifying prompt, and MIMIC-D (Dong et al., 2025) jointly trains decentralised diffusion policies that coordinate implicitly from local information. Both deliberately avoid communication between agents at inference time; neither asks when such a channel becomes necessary, which is the question we study. LLM-based planners such as RoCo (Mandi et al., 2024) and COHERENT (Liu et al., 2025a) employ multi-agent text dialogues for task assignment, but their multi-second query latencies cannot sustain reactive physical handovers. Parameter-Efficient Fine-Tuning (PEFT). PEFT methods such as adapters (Houlsby et al., 2019), LoRA (Hu et al., 2022), and prompt tuning (Lester et al., 2021) adapt a single foundation model to a new domain, and in robotics are used to adapt VLAs to new embodiments or sensors (Kim et al., 2024; Kim et al., 2025). CoLA instead repurposes adapters for communication between agents. 3 Problem Formulation & Methodology 3.1 Decentralised Multi-Agent Formalisation We model multi-robot cooperative manipulation as a decentralised partially observable process with communication, following the Dec-POMDP-Com (Goldman and Zilberstein, 2003) but without a reward function, as is standard in imitation learning (Abbeel and Ng, 2004). The model is defined by the tuple ⟨ℐ,,i,Σ,,i,Ω,ρ0,T⟩ I,S,\A_i\, ,P,\O_i\, , _0,T . Here ℐ=1,…,NI=\1,…,N\ is the set of N robot agents; S is the global state space; iA_i is the continuous action space of agent i; and Σ=ℝdm =R^d_m is the space of messages agents may exchange. The transition distribution (′∣,1,…,N)P( s s, a_1,…, a_N) depends on physical actions only, so messages influence the environment solely through the actions they induce. Each agent i receives observations from its local observation space iO_i, drawn jointly according to Ω(1,…,N∣) ( o_1,…, o_N s). Unlike the standard formulation, which conditions observations on the successor state and joint action, we condition on the current state only, since each observation is a camera rendering of that state. Finally, ρ0 _0 is the initial state distribution and T is the episode horizon. Under decentralised execution, each agent receives only its local observation it∈i o_i^t∈O_i (an RGB camera stream), without direct access to partner observations j≠it o_j≠ i^t or the global state t s^t (a subset of experiments relaxes this with a shared overhead view; those results are marked explicitly). Agents communicate over a channel graph defined by neighbourhood sets (i)⊆ℐ∖iN(i) I \i\: at each timestep t, agent i broadcasts a message it∈Σ m_i^t∈ and receives jtj∈(i)\ m_j^t\_j∈N(i). We restrict attention to reactive policies, which condition on the current observation and the partners’ current messages rather than on full action–observation histories: πi(it|it,jtj∈(i)), _i ( A_i^t\; |\; o_i^t,\ m_j^t\_j∈N(i) ), (1) where it=(it,it+1,…,it+K−1)∈ℝK×Da A_i^t= ( a_i^t, a_i^t+1,…, a_i^t+K-1 ) ^K× D_a denotes an open-loop chunk of K future actions (Da=7D_a=7). This restriction suffices when task-relevant state is either currently observable to some agent or recoverable from current proprioception. Rather than optimising a reward, we are given demonstrations from a scripted expert team policy πE:→1×⋯×Nπ^E:S→A_1×…×A_N, which maps the global state to a joint action (1t,…,Nt)( a_1^t,…, a_N^t) for every agent at each timestep, so each demonstration trajectory records all N agents’ actions synchronously. The expert acts on the global state t s^t, so it is not itself a decentralised reactive policy; a learner can reproduce its behaviour only to the extent that the required information is available through the agent’s own observation or the messages it receives. Our goal is to learn decentralised reactive policies that imitate πEπ^E, which we do by behavioural cloning (section 3.5). 3.2 Frozen Foundation Backbones Each agent i has a pretrained VLA backbone fθf_θ whose parameters remain frozen throughout training. At each timestep, the backbone encodes each of the agent’s ViV_i camera views v∈1,…,Viv∈\1,…,V_i\, together with the task instruction l, into a feature vector, it,v=fθ(it,v,)∈ℝd f_i^t,v=f_θ( o_i^t,v, l) ^d. The agent’s proprioceptive state it q_i^t is mapped to an embedding it=ϕpropi(it) p_i^t= _prop^i( q_i^t) by a small network trained jointly with the other modules. Because proprioception is easier to fit than the visual and message channels, a policy can shortcut through it and under-use those harder signals (Lu et al., 2026); to discourage this, it q_i^t is zeroed with probability 0.10.1 during training. The agent’s representation concatenates these: it=[it,v1∥…∥it,vVi∥it]∈ℝdh. h_i^t= [\, f_i^t,v_1\,\|\,…\,\|\, f_i^t,v_V_i\,\|\, p_i^t\, ] ^d_h. (2) Both the CoLA message encoder and the action head take it h_i^t as input, so a message can carry information from any of the agent’s views and from its physical state. Because fθf_θ is frozen, its features can be computed once and cached, and the trainable modules can only select and transmit information that fθf_θ already encodes. 3.3 CoLA Architectural Design Each agent i discards its backbone’s native action head and instead hosts three trainable modules (fig. 1): Latent Message Encoder (ϕenci _enc^i). A two-layer MLP with ReLU activations that compresses local backbone representations into a coordination message: it=ϕenci(sg[it])∈ℝdm,dm=64, m_i^t= _enc^i (sg_ f[ h_i^t] ) ^d_m, d_m=64, (3) where sg[⋅]sg_ f[·] denotes a stop-gradient applied only to the backbone-feature components of it h_i^t; the proprioceptive encoder ϕpropi _prop^i receives gradients through the message. Message Decoder (ϕdecj→i _dec^j→ i). One decoder per ordered pair, mapping partner j’s message into i’s feature space, j→it=ϕdecj→i(jt)∈ℝd,∀j∈(i). c_j→ i^t= _dec^j→ i( m_j^t) ^d, ∀ j∈N(i). (4) Decoders are not shared across senders: i learns a distinct mapping for each partner, so that in the three-arm task B reads A’s and C’s messages through separate weights. Coordination Action Head. Concatenates the agent’s own representation with its decoded partner embeddings, it=[it∥j→itj∈(i)]∈ℝdh+|(i)|d, z_i^t= [\, h_i^t\; \|\;\ c_j→ i^t\_j∈N(i)\, ] ^d_h+|N(i)|\,d, (5) which conditions the diffusion head of section 3.5: a 1D temporal U-Net with FiLM conditioning that predicts the noise added to the joint chunk, alongside a logit head for the binary gripper column. Figure 1: CoLA framework overview. (a) Decentralised inference: each agent’s frozen VLA backbone (Octo-Base or π0.5 _0.5) encodes its local observation. Its CoLA module compresses the backbone feature into a compact message (dm=64d_m=64), decodes the partner’s message, and concatenates the result with its own representation to form i z_i. This vector conditions a U-Net action head via FiLM, which predicts a coordinated action chunk (K=10K=10). (b) Adapter and action-head training: backbone features are cached offline and the backbones stay frozen. Only the CoLA modules are trained with a behavioural cloning loss, with gradients flowing end-to-end through the communication channel. (c) Multi-agent broadcast: each agent broadcasts its message to all other agents and receives theirs, so the same design scales to N≥3N≥ 3 agents. Action Chunking Dynamics. Actions are generated in chunks of K=10K=10 steps executed open-loop at 50Hz50\,Hz control rate (0.2s0.2\,s per chunk). The policy is queried once per chunk, giving a 5Hz5\,Hz inference rate (one query every 200ms200\,ms). 3.4 Multi-Agent Topology Routing CoLA extends to N agents through flexible graph topologies: Bimanual Teams (N=2N=2): Direct peer-to-peer exchange, (1)=2N(1)=\2\ and (2)=1N(2)=\1\. Each agent’s head input is [i∥j→i][ h_i\,\|\, c_j→ i], of dimension dh+d_h+d. Tri-Manual System (N=3N=3): A fully connected topology, (i)=ℐ∖iN(i)=I \i\ for every agent, giving six directed channels. Each agent decodes both partners’ messages through its own per-sender decoder, i=[i‖j→i‖k→i] z_i=[ h_i\,\|\, c_j→ i\,\|\, c_k→ i] of dimension dh+2d_h+2d. The architecture is symmetric: no agent acts as a hub, all three heads take identically shaped input, and encoder and decoder architectures are identical across agents. 3.5 End-to-End Imitation Learning Objective Given a multi-agent demonstration dataset =(1:Nt,1:Nt)D=\( o_1:N^t, A_1:N^t)\ collected from scripted experts, CoLA is trained by behavioural cloning. We split each expert chunk as it=[it,~it] A_i^t=[ Q_i^t, g_i^t], where it∈ℝK×6 Q_i^t ^K× 6 holds the joint targets and ~it∈−0.9,0.9K g_i^t∈\-0.9,0.9\^K the normalised gripper channel, from which we derive binary labels it=[~it>0]∈0,1K y_i^t= 1[ g_i^t>0]∈\0,1\^K for the classification loss. Following diffusion policies (Ho et al., 2020; Chi et al., 2025), we sample, independently for each agent, a diffusion step τi _i uniformly from 1,…,S−1\1,…,S-1\ with S=100S=100, and noise ϵi∼(,) ε_i ( 0, I). We then form the corrupted joint chunk ~it=α¯τiit+1−α¯τiϵi Q_i^t= α_ _i\, Q_i^t+ 1- α_ _i\, ε_i under a cosine schedule (Nichol and Dhariwal, 2021), and train each agent’s head ϕuneti _unet^i to predict the added noise conditioned on it z_i^t: ℒ(ϕ)=[∑i=1Nτi,ϵi[ℒi]],ℒi=16K‖ϕuneti(~it,τi,it)−ϵi‖F2+λBCE(ϕgripi(it),it),L( φ)=E_D [ _i=1^NE_ _i,\, ε_i [L_i ] ], L_i= 16K \| _unet^i( Q_i^t, _i, z_i^t)- ε_i \|_F^2+λ\,BCE ( _grip^i( z_i^t),\, y_i^t ), (6) where ϕgripi _grip^i outputs K gripper logits, the BCE is averaged over the K steps, and λ=0.25λ=0.25. At test time, we sample with DDIM (Song et al., 2020) using 1010 denoising steps. We predict the gripper with a separate classification head because it is binary: in early development, predicting it as a continuous dimension produced near-constant outputs, which motivated the separate head. The classification head reaches 98.798.7–100%100\% validation gripper accuracy across all four coordination regimes (table 14). Gradients Through the Channel. Because it z_i^t depends on partners’ messages, gradients flow across the channel. Agent j’s message jt m_j^t enters only its receivers’ action heads, so j’s encoder is trained entirely by its partners’ losses, whereas the decoder producing j→it c_j→ i^t belongs to agent i and is trained by ℒiL_i alone. Each encoder is therefore pushed to transmit whatever its partners need to disambiguate their actions. Offline Feature Caching. Because fθf_θ is frozen, its features are computed once per dataset and reused by every run and ablation, so training CoLA takes under one GPU-hour, against 22–4949 for the fine-tuned baselines at the same optimiser budget (table 5). 4 Experiments Our experiments aim to answer the following questions: (Q1) Can agents built on frozen VLA backbones coordinate under decentralised execution when only a message channel and a new action head are trained? (Q2) How much does the message channel contribute to success? (Q3) Does the approach generalise to three agents and to different frozen backbones? 4.1 Tasks We evaluate on three tasks that differ in how much task-relevant information each agent can observe with its own cameras (fig. 2). Figure 2: Multi-Agent Manipulation Progression. Key milestones across evaluated tasks: (top) Occluded Tray-Handover Task, (middle) 2-arm Reactive Handover, and (bottom) 3-arm Handover Chain. Occluded tray handover. Arm A grasps a box bearing one of three coloured markers (blue, green, or yellow) on the face pointing toward Arm A, and hands it to Arm B, which must place it on the tray of the matching colour on its own side of the table. Each arm observes only its own wrist camera. The marker appears in Arm A’s camera in 100%100\% of control steps and in Arm B’s in under 5%5\%. Without communication, Arm B therefore has almost no access to the colour, and among the episodes in which it places the box it can expect to select the correct tray only about one time in three. An episode is successful if the box enters the correct tray. Two-arm handover. Arm A starts holding the object and presents it at a handoff point; Arm B must approach, grasp, and hold while A releases. A’s presentation pose is randomised by ±10cm± 10\,cm laterally and vertically. The challenge is timing: B must close its gripper before A opens its own, and a premature release drops the object. We evaluate two variants: each arm’s wrist camera only, and each arm’s wrist camera together with a shared overhead camera. An episode is successful if Arm B grasps the object without dropping it. Three-arm handover chain. The object passes A → B → C, with A’s presentation randomised as above, and all arms additionally share the overhead view alongside their own wrist camera. B must receive the object and pass it on, C’s approach depends on when B is ready, and a failure at either transfer ends the episode. An episode is successful if both transfers complete and Arm C then holds the object for 2020 steps. 4.2 Experimental Setup Simulation Environments. All environments are simulated in MuJoCo (Todorov et al., 2012) with a 2ms2\,ms physics timestep (500Hz500\,Hz) and a control decimation of 1010, giving a 50Hz50\,Hz control rate. We control N∈2,3N∈\2,3\ 6-DoF Trossen ViperX 300 arms with parallel-jaw grippers. Each arm’s action is 77-dimensional: target positions for its six revolute joints, normalised per dimension to [−0.9,0.9][-0.9,0.9] using min–max statistics from the training split, plus a gripper command, which CoLA predicts as a binary open/closed decision. Each agent observes a 256×256256× 256 RGB image from its own wrist-mounted camera, together with its 77-dimensional proprioceptive state (six joint angles and gripper opening). In the overhead variants of the two- and three-arm handover, every agent additionally observes a shared fixed overhead camera alongside its own wrist camera (section 3.1), for CoLA and the baselines alike. The occluded tray-handover task uses wrist cameras only. All three tasks use Octo-Base (9393M parameters). On the occluded tray-handover task we additionally evaluate π0.5 _0.5 (base checkpoint, ∼3.3 3.3B parameters (Physical Intelligence et al., 2025)). Both backbones are frozen in their entirety and run once offline, encoding each agent’s observation at each timestep into a single cached feature vector: the 768768-dimensional readout embedding for Octo-Base and, because π0.5 _0.5 has no readout token, a padding-masked mean over its final-layer PaliGemma prefix tokens (image, prompt, and joint state as 256256-bin tokens, matching the LoRA baseline) giving a 20482048-dimensional vector. On this backbone, the 77-d proprioceptive state therefore enters CoLA twice: once inside the frozen 20482048-D feature (via the prompt tokens) and again through ϕprop _prop (section 3.2); the proprio dropout described there regularises only the latter copy. Neither backbone’s own action head is used. CoLA’s modules (section 3.3) have the same architecture across backbones; only the layer widths that depend on the backbone feature dimension change (14.514.5M trainable parameters with Octo-Base, 16.916.9M with π0.5 _0.5, of which the communication channel accounts for 0.330.33M and 0.830.83M respectively and the remainder is the per-arm action head and proprioceptive encoder). CoLA uses a single generic instruction, “coordinate with partner”, for every arm and task, whereas the baselines receive role-specific instructions (table 4). No instruction names the marker colour. Pretrained Octo does not complete the two-arm handover zero-shot with its own action head (table 13). Evaluation Protocol. For the handover tasks, each policy is evaluated on 200200-episode blocks at evaluation-seed offsets drawn from 00, 10001000, 20002000, 30003000 and 40004000, and we report the mean ± standard deviation of success rate across the blocks evaluated. The number of blocks differs by cell: five for every method on the 2-arm overhead task except CoLA with zeroed messages (three); three for every method on the 2-arm wrist-only task (offsets 00, 10001000, 20002000); and five for the baselines but a single block (offset 00) for both CoLA rows on the three-arm chain. The per-seed blocks behind every cell are given in tables 6, 7 and 8. The occluded tray-handover task keeps the handover positions fixed and randomises only hidden marker colour among three discrete values. A single seed with 150150 episodes already samples each colour densely (∼50 50 episodes per colour), therefore this task is evaluated with only a single seed (00). For this task we report success rate with a 95%95\% Wilson confidence interval over episodes, with the per-stage breakdown in table 10. Baselines for comparison. All baselines are trained on the same demonstrations and episode splits as CoLA and, unless stated otherwise, on the same Octo-Base backbone. Within each task, every method uses the same number of optimiser steps at the same effective batch size of 128128 (12,60012,600, 20,25020,250 and 26,85026,850 steps for the two-arm handover, tray and three-arm handover tasks, respectively). This matches the optimisation budget, not compute or trainable capacity. Learning rates and optimisers follow each method’s standard recipe (Adam without weight decay for CoLA; AdamW with weight decay 10−410^-4 for the Octo baselines), and full architectural and optimisation settings are listed in table 3 (appendix B). Since CoLA trains on cached backbone features, matching optimiser steps gives the fine-tuned baselines substantially more compute than CoLA. • Independent decentralised policies: one policy per arm, conditioned on its wrist camera on the wrist-only tasks, or on both its wrist camera and the shared overhead camera on the overhead tasks, trained either (a) head-only, with the backbone frozen and a freshly initialised action head, or (b) with full fine-tuning of all backbone weights on the multi-robot demonstrations. All Octo baselines receive proprioception through Octo’s proprio tokenizer: each decentralised policy sees its own arm’s 7-D joint state, and the centralised policy both arms’ (14-D). Because Octo-Base was pretrained without proprioception, this projection is newly initialised; head-only baselines keep it frozen at that initialisation, while full fine-tuning trains it. • Centralised policy: a single model that outputs the joint 14-D action, in the same two regimes, from both arms’ wrist cameras on the wrist-only tasks, or from both arms’ wrist cameras plus the shared overhead camera on the overhead tasks. It violates decentralised execution and serves as a privileged reference. • CoLA with zeroed messages: the full CoLA architecture, trained and evaluated with all messages set to zero. It has the same architecture, parameter count, and training recipe as CoLA; the only difference is that the channel carries no information, which isolates the contribution of communication. • Cross-backbone transfer (occluded tray-handover task): independent per-arm π0.5 _0.5 policies fine-tuned with openpi’s LoRA recipe (LoRA on both Gemma experts plus the full vision encoder; 467.0467.0M trainable parameters per arm), compared against CoLA on the frozen π0.5 _0.5 backbone (8.58.5M per arm), testing whether CoLA depends on a particular backbone. 4.3 Main Empirical Results Table 1: Main results. Success rate (%), ordered from least to most shared observability. HO = handover. Handover columns: mean ± std over 200-episode evaluation blocks (section 4.2); every row is a single training run. Tray: single 150-episode block. –: not evaluated. Method Occluded Tray-HO (wrist only) 2-Arm HO (wrist only) 2-Arm HO (overhead) 3-Arm HO (overhead) CoLA (with message) 76.0 74.8 ± 2.6 90.6 ± 1.7 63.5† CoLA (message zeroed) 1.3 64.3 ± 6.7 88.3 ± 2.9 60.0† Decentralised, head-only 22.7 43.8 ± 2.0 68.3 ± 2.6 20.5 ± 2.7 Centralised, head-only 16.0 21.0 ± 2.2 65.9 ± 4.7 — Centralised, full fine-tune 10.7 80.0 ± 2.6 92.1 ± 2.0 — Decentralised, full fine-tune 12.0 69.3 ± 1.8 82.4 ± 3.1 31.6 ± 4.5 † single 200-episode block. Tray chance is 33.3% conditional on reaching a tray. Per-seed and per-stage results: appendix D. Q1: Coordination Efficacy. table 1 shows that CoLA outperforms every decentralised baseline on all four tasks, while modifying only lightweight adapters and an action head atop a frozen backbone. On the occluded tray-handover task, independent decentralised policies perform poorly, while CoLA reaches 76.0%76.0\%. On the handover tasks, CoLA exceeds the decentralised head-only baseline by 31.031.0 percentage points on wrist-only handover (74.8%74.8\% vs. 43.8%43.8\%), 22.322.3 percentage points with a shared overhead view (90.6%90.6\% vs. 68.3%68.3\%), and 43.043.0 points at three arms (63.5%63.5\% vs. 20.5%20.5\%). CoLA also approaches the privileged centralised full fine-tune on handover (90.6%90.6\% vs. 92.1%92.1\% with a shared view; 74.8%74.8\% vs. 80.0%80.0\% wrist-only) while retaining fully decentralised execution. fig. 3 contrasts a CoLA rollout with an uncoordinated baseline, and appendix F categorises the failures. Q2: Is Communication Causally Necessary? The head-only baseline underperforms no-message CoLA on every handover task (e.g., 68.3%68.3\% vs. 88.3%88.3\% with a shared overhead view). Additionally, CoLA replaces Octo’s action head with a diffusion U-Net with FiLM conditioning, predicts the gripper with a separate classification head, adds a learned proprioceptive embedding, and uses a different chunk length and training schedule. Comparisons against the Octo baselines therefore measure CoLA’s whole pipeline, not communication. The no-message ablation isolates the channel: it keeps CoLA’s architecture and training recipe unchanged and removes only the message content. Removing messages collapses the occluded tray handover (76.0%76.0\% to 1.3%1.3\%), although most of this drop occurs before tray selection: the message-zeroed model completes the handover itself in only 8.0%8.0\% of episodes, against 88.7%88.7\% for the head-only baseline, which also has no channel (table 10). Training with a constant zero message therefore impairs the whole policy on this task, so 1.3%1.3\% is not just the cost of withholding the colour alone. The evidence that the channel carries the colour is tray choice given a placement (85.7%85.7\% for CoLA against 26.4%26.4\% for the head-only baseline, chance 33.3%33.3\%) and the probes of section 4.4. Removing messages also degrades the wrist-only handover (74.8%74.8\% to 64.3%64.3\%), but has no measurable effect when a shared overhead view shows each agent its partner (90.6%90.6\% vs. 88.3%88.3\% at two arms; 63.5%63.5\% vs. 60.0%60.0\% at three, within noise). The channel therefore matters to the extent that task-relevant information is private to one agent. When agents share a view, CoLA’s advantage over the independent baselines comes from the pipeline differences above, which we do not isolate individually. table 11 shows that performance is largely insensitive to message width dm∈16,64,128d_m∈\16,64,128\. Table 2: Cross-backbone transfer on the occluded tray-handover task. CoLA on a frozen π0.5 _0.5 backbone versus independently LoRA-fine-tuned per-arm π0.5 _0.5 policies. Placed: the box reaches any tray. Correct ∣ placed: the box reaches the correct tray, given that it reaches one (chance 33.3%33.3\%). Overall: the box reaches the correct tray, over all episodes. Brackets give 95% Wilson intervals; bold marks a value whose interval does not overlap the other method’s. Method Params (per arm) Train (GPU-h) Placed (%) Correct ∣ placed (%) Overall (%) CoLA 8.58.5M 0.20.2 52.0 [44.1, 59.8] 71.8 [61.0, 80.6] 37.3 [30.0, 45.3] Decentralised, LoRA fine-tune 467.0467.0M 24.424.4 100.0 [97.5, 100] 34.7 [27.5, 42.6] 34.7 [27.5, 42.6] Q3: Does CoLA’s Advantage Persist with Team Size and Backbone? At three arms, CoLA outperforms decentralised full fine-tuning by 31.931.9 percentage points (63.5%63.5\% vs. 31.6%31.6\%), up from 8.28.2 percentage points at two arms (90.6%90.6\% vs. 82.4%82.4\%), and the head-only baseline by 43.043.0 points (up from 22.322.3). Because zeroed-message CoLA reaches 60%60\%, this widening margin reflects CoLA’s architecture rather than communication when an overhead view is shared; whether the channel’s own contribution grows with team size requires a three-arm task with private information. Per-stage transfer rates for the chain are reported in table 9. To test whether the channel’s benefit is specific to Octo, we repeat the occluded tray-handover task with a frozen π0.5 _0.5 backbone (table 2). The two methods fail in different places. The LoRA-fine-tuned per-arm policies place the box on a tray in every episode but choose the correct tray at chance level (34.7%34.7\% against a chance rate of 33.3%33.3\%), consistent with the placing arm having no access to the hidden colour. CoLA places the box in only 52.0%52.0\% of episodes, but when it does, it selects the correct tray 71.8%71.8\% of the time (95% CI [61.0,80.6][61.0,80.6]), well above both chance and the baseline. The message channel therefore still carries the hidden task information on π0.5 _0.5, what does not transfer is manipulation reliability, and overall success is statistically indistinguishable (37.3%37.3\% vs. 34.7%34.7\%). 4.4 Probing What the Channel Encodes We analyse the information transmitted in the message by fitting linear probes (multinomial logistic regression for colour, ridge regression for pose) from three representations of the sending agent: its own features it h_i^t (self), its message it m_i^t from the trained checkpoint (message), and the encoder output from the message-zeroed checkpoint (control), which receives no gradient when messages are zeroed (section 3.5) and so stays at random initialisation. On the occluded tray-handover task, the message decodes occluded box colour at 73.3%73.3\% versus chance-level 33.3%33.3\% for the control (episode-level majority vote; per frame, 44.2%44.2\% vs. 39.7%39.7\%), a 4040-point gap attributable to training through the channel that corroborates the necessity results (fig. 4 and table 12, appendix D). 5 Conclusion and Limitations Conclusion. We presented CoLA, a framework that trains a message-passing channel and a new action-generation pipeline on top of frozen, independently deployed VLA backbones, and used it to measure when that channel is actually necessary for coordination. Isolating the channel’s effect from the rest of the architecture, we found communication is not uniformly beneficial. It is necessary when task-relevant state is private to one agent, and has no measurable effect once agents share a view of the workspace. On the occluded tray-handover task, the channel also carried the hidden colour with a frozen π0.5 _0.5 backbone, although overall success there did not exceed that of the LoRA-fine-tuned baseline. Our results suggest that explicit communication and implicit, shared-context coordination are not competing solutions but complementary ones, each suited to a different information regime, and that determining which regime a task falls into is itself a necessary step in designing multi-agent VLA systems. Limitations. All experiments run in simulation on scripted expert demonstrations, and we have not tested transfer to physical robots. CoLA also assumes zero-latency synchronous communication. Although realistic for local high-speed bus architectures, performance under asynchronous network dropouts and variable latency remains to be benchmarked. Additionally, CoLA connects agents with a fully connected topology and a separate decoder for every ordered pair (section 3.4), so both the number of channels and the number of decoders grow as O(N2)O(N^2), with six of each at N=3N=3. This is practical at the team sizes we study, but scaling to dozens of robots will require sharing decoder weights across senders, or aggregating messages through graph neural network or cross-attention routers (Das et al., 2019). Reproducibility Statement Environments use MuJoCo 3.4.0 with the Trossen ViperX 300s models from MuJoCo Menagerie (physics 500 Hz, control 50 Hz). All demonstrations are successful expert episodes, split 80/10/10 by episode. Frozen Octo-Base features (and Octo-Small, used only in table 13) are pre-extracted in JAX and cached. Hyperparameters are listed in table 3 and the full procedure in algorithm 1. Experiments ran on one node with 8 NVIDIA A100 (40 GB) GPUs. Code: https://github.com/Ahaskar04/CoLA.git. AI Use Statement Generative AI tools were used exclusively for rephrasing and condensing author written text, formatting assistance, spelling and grammar refinement. All conceptual formulations, experimental analyses, and technical claims were designed and verified by the authors. References Abbeel and Ng (2004) P. Abbeel and A. Y. Ng Apprenticeship Learning via Inverse Reinforcement Learning. In Proceedings of the International Conference on Machine Learning (ICML), p. 1. Cited by: §3.1. Albrecht et al. (2024) S. V. Albrecht, F. Christianos, and L. Schäfer Multi-Agent Reinforcement Learning: Foundations and Modern Approaches. MIT Press. Cited by: §2. Asmar and Kochenderfer (2024) D. M. Asmar and M. J. Kochenderfer Efficient Multiagent Planning via Shared Action Suggestions. arXiv preprint. External Links: Link Cited by: §2. Baker et al. (2020) B. Baker, I. Kanitscheider, T. Markov, Y. Wu, G. Powell, B. McGrew, and I. Mordatch Emergent Tool Use From Multi-Agent Autocurricula. In Proceedings of the International Conference on Learning Representations (ICLR), Cited by: §2. Brohan et al. (2023a) A. Brohan, N. Brown, J. Carbajal, Y. Chebotar, X. Chen, K. Choromanski, T. Ding, D. Driess, A. Dubey, C. Finn, P. Florence, C. Fu, M. Gonzalez Arenas, K. Gopalakrishnan, K. Han, K. Hausman, A. Herzog, J. Hsu, B. Ichter, A. Irpan, N. Joshi, R. Julian, D. Kalashnikov, Y. Kuang, I. Leal, L. Lee, T. E. Lee, S. Levine, Y. Lu, H. Michalewski, I. Mordatch, K. Pertsch, K. Rao, K. Reymann, M. Ryoo, G. Salazar, P. Sanketi, P. Sermanet, J. Singh, A. Singh, R. Soricut, H. Tran, V. Vanhoucke, Q. Vuong, A. Wahid, S. Welker, P. Wohlhart, J. Wu, F. Xia, T. Xiao, P. Xu, S. Xu, T. Yu, and B. Zitkovich RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control. In Proceedings of the Conference on Robot Learning (CoRL), Cited by: §1. Brohan et al. (2023b) A. Brohan, N. Brown, J. Carbajal, Y. Chebotar, J. Dabis, C. Finn, K. Gopalakrishnan, K. Hausman, A. Herzog, J. Hsu, J. Ibarz, B. Ichter, A. Irpan, T. Jackson, S. Jesmonth, N. J. Joshi, R. Julian, D. Kalashnikov, Y. Kuang, I. Leal, K. Lee, S. Levine, Y. Lu, U. Malla, D. Manjunath, I. Mordatch, O. Nachum, C. Parada, J. Peralta, E. Perez, K. Pertsch, J. Quiambao, K. Rao, M. Ryoo, G. Salazar, P. Sanketi, K. Sayed, J. Singh, S. Sontakke, A. Stone, C. Tan, H. Tran, V. Vanhoucke, S. Vega, Q. Vuong, F. Xia, T. Xiao, P. Xu, S. Xu, T. Yu, and B. Zitkovich RT-1: Robotics Transformer for Real-World Control at Scale. In Proceedings of Robotics: Science and Systems (RSS), Cited by: §1. Chi et al. (2025) C. Chi, Z. Xu, S. Feng, E. Cousineau, Y. Du, B. Burchfiel, R. Tedrake, and S. Song Diffusion Policy: Visuomotor Policy Learning via Action Diffusion. The International Journal of Robotics Research. Cited by: §3.5. Das et al. (2019) A. Das, T. Gervet, J. Romoff, D. Batra, D. Parikh, M. Rabbat, and J. Pineau TarMAC: Targeted Multi-Agent Communication. In Proceedings of the International Conference on Machine Learning (ICML), Cited by: §2, §5. Dey et al. (2026) S. Dey, R. Riener, S. Dosen, and S. V. Albrecht From Autonomy to Alliance: Robotic Foundation Models Must Learn with Us, Not Just for Us. Science Robotics 11 (113), p. eaea1822. Cited by: §1. Dong et al. (2025) D. Dong, M. Bhatt, S. Choi, and N. Mehr MIMIC-D: Multi-modal Imitation for Multi-agent Coordination with Decentralized Diffusion Policies. arXiv preprint arXiv:2509.14159. Cited by: §2. Doshi et al. (2026) R. Doshi, T. Gao, A. Chen, C. Finn, and J. Bohg CHORUS: Decentralized Multi-Embodiment Collaboration with One VLA Policy. External Links: 2606.12352, Link Cited by: §2. Foerster et al. (2016) J. Foerster, Y. M. Assael, N. de Freitas, and S. Whiteson Learning to Communicate with Deep Multi-Agent Reinforcement Learning. In Advances in Neural Information Processing Systems (NeurIPS), Cited by: §2. Fu et al. (2024) Z. Fu, T. Z. Zhao, and C. Finn Mobile ALOHA: Learning Bimanual Mobile Manipulation with Low-Cost Whole-Body Teleoperation. In Proceedings of the Conference on Robot Learning (CoRL), Cited by: §1, §2. Goldman and Zilberstein (2003) C. V. Goldman and S. Zilberstein Optimizing Information Exchange in Cooperative Multi-Agent Systems. In Proceedings of the International Conference on Autonomous Agents and Multi-Agent Systems (AAMAS), Cited by: §3.1. Havrylov and Titov (2017) S. Havrylov and I. Titov Emergence of Language with Multi-agent Games: Learning to Communicate with Sequences of Symbols. In Advances in Neural Information Processing Systems (NeurIPS), Cited by: §2. Ho et al. (2020) J. Ho, A. Jain, and P. Abbeel Denoising Diffusion Probabilistic Models. In Advances in Neural Information Processing Systems (NeurIPS), Vol. 33, p. 6840–6851. Cited by: §3.5. Houlsby et al. (2019) N. Houlsby, A. Giurgiu, S. Jastrzebski, B. Morrone, Q. de Laroussilhe, A. Gesmundo, M. Attariyan, and S. Gelly Parameter-Efficient Transfer Learning for NLP. In Proceedings of the International Conference on Machine Learning (ICML), Cited by: §2. Hu et al. (2022) E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, and W. Chen LoRA: Low-Rank Adaptation of Large Language Models. In Proceedings of the International Conference on Learning Representations (ICLR), Cited by: §2. Im et al. (2026) H. Im, E. Jeong, A. Kolobov, J. Fu, and Y. Lee TwinVLA: Data-Efficient Bimanual Manipulation with Twin Single-Arm Vision-Language-Action Models. In Proceedings of the International Conference on Learning Representations (ICLR), Cited by: §2. Kim et al. (2025) M. J. Kim, C. Finn, and P. Liang Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success. In Proceedings of Robotics: Science and Systems (RSS), Cited by: §2. Kim et al. (2024) M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. Foster, G. Lam, P. Sanketi, Q. Vuong, T. Kollar, B. Burchfiel, R. Tedrake, D. Sadigh, S. Levine, P. Liang, and C. Finn OpenVLA: An Open-Source Vision-Language-Action Model. In Proceedings of the Conference on Robot Learning (CoRL), Cited by: §1, §2. Krnjaic et al. (2024) A. Krnjaic, R. D. Steleac, J. D. Thomas, G. Papoudakis, L. Schäfer, A. W. K. To, K. Lao, M. Cubuktepe, M. Haley, P. Börsting, and S. V. Albrecht Scalable Multi-Agent Reinforcement Learning for Warehouse Logistics with Robotic and Human Co-Workers. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Cited by: §1. Lazaridou et al. (2017) A. Lazaridou, A. Peysakhovich, and M. Baroni Multi-Agent Cooperation and the Emergence of (Natural) Language. In Proceedings of the International Conference on Learning Representations (ICLR), Cited by: §2. Lee et al. (2024) A. Lee, I. Chuang, L. Chen, and I. Soltani InterACT: Inter-dependency Aware Action Chunking with Hierarchical Attention Transformers for Bimanual Manipulation. In Proceedings of the Conference on Robot Learning (CoRL), Cited by: §2. Lester et al. (2021) B. Lester, R. Al-Rfou, and N. Constant The Power of Scale for Parameter-Efficient Prompt Tuning. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), Cited by: §2. Liu et al. (2025a) K. Liu, Z. Tang, D. Wang, Z. Wang, X. Li, and B. Zhao COHERENT: Collaboration of Heterogeneous Multi-Robot System with Large Language Models. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), Cited by: §1, §2. Liu et al. (2025b) S. Liu, L. Wu, B. Li, H. Tan, H. Chen, Z. Wang, K. Xu, H. Su, and J. Zhu RDT-1B: a Diffusion Foundation Model for Bimanual Manipulation. In Proceedings of the International Conference on Learning Representations (ICLR), Cited by: §1, §2. Lu et al. (2025) G. Lu, T. Yu, H. Deng, S. S. Chen, Y. Tang, and Z. Wang AnyBimanual: Transferring Unimanual Policy for General Bimanual Manipulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), Cited by: §2. Lu et al. (2026) Y. Lu, Y. Tian, Z. Yuan, X. Wang, P. Hua, Z. Xue, and H. Xu H3^3DP: Triply-Hierarchical Diffusion Policy for Visuomotor Learning. In Proceedings of the International Conference on Learning Representations (ICLR), Cited by: §3.2. Mandi et al. (2024) Z. Mandi, S. Jain, and S. Song RoCo: Dialectic Multi-Robot Collaboration with Large Language Models. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), Cited by: §1, §2. Nichol and Dhariwal (2021) A. Q. Nichol and P. Dhariwal Improved Denoising Diffusion Probabilistic Models. In Proceedings of the International Conference on Machine Learning (ICML), Vol. 139, p. 8162–8171. Cited by: §3.5. Octo Model Team et al. (2024) Octo Model Team, D. Ghosh, H. Walke, K. Pertsch, K. Black, O. Mees, S. Dasari, J. Hejna, T. Kreiman, C. Xu, J. Luo, Y. L. Tan, L. Y. Chen, P. Sanketi, Q. Vuong, T. Xiao, D. Sadigh, C. Finn, and S. Levine Octo: An Open-Source Generalist Robot Policy. In Proceedings of Robotics: Science and Systems (RSS), Cited by: §1. Physical Intelligence et al. (2025) Physical Intelligence, K. Black, N. Brown, J. Darpinian, K. Dhabalia, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, M. Y. Galliker, D. Ghosh, L. Groom, K. Hausman, B. Ichter, S. Jakubczak, T. Jones, L. Ke, D. LeBlanc, S. Levine, A. Li-Bell, M. Mothukuri, S. Nair, K. Pertsch, A. Z. Ren, L. X. Shi, L. Smith, J. T. Springenberg, K. Stachowicz, J. Tanner, Q. Vuong, H. Walke, A. Walling, H. Wang, L. Yu, and U. Zhilinsky π0.5 _0.5: a Vision-Language-Action Model with Open-World Generalization. arXiv preprint arXiv:2504.16054. Cited by: §4.2. Song et al. (2020) J. Song, C. Meng, and S. Ermon Denoising Diffusion Implicit Models. arXiv preprint arXiv:2010.02502. Cited by: §3.5. Sukhbaatar et al. (2016) S. Sukhbaatar, A. Szlam, and R. Fergus Learning Multiagent Communication with Backpropagation. In Advances in Neural Information Processing Systems (NeurIPS), Cited by: §2. Todorov et al. (2012) E. Todorov, T. Erez, and Y. Tassa MuJoCo: A Physics Engine for Model-Based Control. In Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Cited by: §4.2. Yifan et al. (2025) C. Yifan, M. Wei, X. Wang, Y. Liu, J. Wang, H. Song, L. Ma, D. Di, C. Sun, K. Liu, L. Qi, J. Yu, X. Tian, S. Liang, C. Duan, Z. Hong, W. Zhang, and T. Liu Embodied AI: A Survey on the Evolution from Perceptive to Behavioral Intelligence. SmartBot 1 (3), p. e70003. Cited by: §1. Zhao et al. (2023) T. Z. Zhao, V. Kumar, S. Levine, and C. Finn Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware. In Proceedings of Robotics: Science and Systems (RSS), Cited by: §1, §2. Appendix A Appendix A Training Algorithm B Appendix B Detailed Architectural Parameters & Training Budget B.1 B.1 CoLA Specifications and Hyperparameters B.2 B.2 Language Instructions B.3 B.3 Trainable Capacity and Training Cost C Appendix C Task Suite C.1 C.1 Qualitative Trajectory Comparison D Appendix D Full Experiment Results D.1 D.1 Two-Arm Handover Results D.2 D.2 Three-Arm Handover Chain Results D.3 D.3 Occluded Tray-Handover Results D.4 D.4 Linear Probes of Channel Content D.5 D.5 Zero-Shot Pretrained Octo D.6 D.6 Validation Gripper Accuracy E Appendix E Additional Training Dynamics F Appendix F Failure Mode Taxonomy F.1 F.1 Upstream Grasp Failures F.2 F.2 Temporal Desynchronisation Failures F.3 F.3 Post-Transfer Release Failures Appendix A Training Algorithm Algorithm 1 CoLA Optimisation with Offline Feature Caching (diffusion U-Net head) Multi-agent demonstration dataset =(1:Nt,1:Nt,1:Nt)t=1TD=\( o_1:N^t, q_1:N^t, a_1:N^t)\_t=1^T (images, proprioception, actions); frozen VLA backbone fθf_θ with instruction l; chunk horizon K=10K=10; message dimension dm=64d_m=64; diffusion steps S=100S=100 with cosine schedule α¯τ=0S−1\ α_τ\_τ=0^S-1; gripper weight λ=0.25λ=0.25; proprio dropout p=0.1p=0.1. // Phase 1: One-Time Offline Feature Extraction Split D by episode into train,val,testD_train,D_val,D_test (80/10/10). for each split s∈train,val,testD_s∈\D_train,D_val,D_test\ do for each batch of multi-agent observations 1:N o_1:N do i←sg[fθ(i,).readout_action[:,−1,0,:]],∀i∈1,…,N f_i [f_θ( o_i, l). readout\_action[:,-1,0,:] ], ∀ i∈\1,…,N\ Cache (1:N,1:N,1:N)( f_1:N, q_1:N, a_1:N) to disk. end for end for // Phase 2: Action Chunk Indexing Index valid starting timesteps =t∣[t,t+K−1] belongs to the same episodeV=\t [t,t+K-1] belongs to the same episode\. // Phase 3: Adapter Parameter Optimisation Initialise ϕ=ϕpropi,ϕenci,ϕdecj→ij∈(i),ϕuneti,ϕgripii=1N φ=\ _prop^i, _enc^i,\ _dec^j→ i\_j∈N(i), _unet^i, _grip^i\_i=1^N; best_val_loss←∞best\_val\_loss←∞. for epoch =1=1 to 150150 do for minibatch ℬ⊂trainB⊂V_train of size 128 do Load cached it,iti=1N\ f_i^t, q_i^t\_i=1^N and target chunks it=it:t+K−1i=1N\ A_i^t= a_i^t:t+K-1\_i=1^N. for agent i=1i=1 to N do Drop proprioception: ~it←biit,bi∼Bernoulli(1−p) q_i^t← b_i\, q_i^t,\;b_i (1-p). Self representation: it←[sg[it]∥ϕpropi(~it)] h_i^t←[sg[ f_i^t]\,\|\, _prop^i( q_i^t)]. Encode message: it←ϕenci(it) m_i^t← _enc^i( h_i^t). end for for agent i=1i=1 to N do Decode partner messages: j→it←ϕdecj→i(jt),∀j∈(i) c_j→ i^t← _dec^j→ i( m_j^t),\;∀ j∈N(i). Conditioning vector: it←[it∥j→itj∈(i)] z_i^t←[ h_i^t\,\|\,\ c_j→ i^t\_j∈N(i)]. Split target: joint chunk ijnt∈ℝK×6 A_i^jnt ^K× 6, gripper labels i←[~it>0]∈0,1K y_i 1[ g_i^t>0]∈\0,1\^K. Sample τi∼1,…,S−1 _i \1,…,S-1\ and ϵi∼(,) ε_i (0, I), independently per agent. Noise the joints: ~i←α¯τiijnt+1−α¯τiϵi A_i← α_ _i\, A_i^jnt+ 1- α_ _i\, ε_i. Predict noise: ϵ^i←ϕuneti(~i,τi,it) ε_i← _unet^i( A_i, _i, z_i^t) ⊳ 1D temporal U-Net, FiLM on MLP([emb(τi)∥it])MLP([emb( _i)\| z_i^t]) Predict gripper logits: i←ϕgripi(it)∈ℝK g_i← _grip^i( z_i^t) ^K. end for Compute joint loss: ℒ(ϕ)←∑i=1N(16K‖ϵ^i−ϵi‖22+λ1K∑k=1KBCE(gi,k,yi,k))L( φ)← _i=1^N ( 16K\,\| ε_i- ε_i\|_2^2+λ\, 1K _k=1^KBCE(g_i,k,y_i,k) ), averaged over ℬB. Update ϕ←Adam(ϕ,∇ϕℒ) φ ( φ, _ φL) with gradient clipping ‖g‖2≤1.0\|g\|_2≤ 1.0 and cosine learning-rate decay 10−4→10−510^-4→ 10^-5. end for Evaluate ℒvalL_val on valV_val (no proprio dropout; fixed τ,ϵτ, ε draws). if ℒval<best_val_lossL_val<best\_val\_loss then Save optimal parameters ϕ∗←ϕ φ^*← φ; best_val_loss←ℒvalbest\_val\_loss←L_val. else if epochs without validation improvement >25>25 then break (Early stopping triggered). end if end for return ϕ∗ φ^* Appendix B Detailed Architectural Parameters & Training Budget B.1 CoLA Specifications and Hyperparameters Table 3: CoLA specifications and hyperparameters. Parameter counts are exact, read from the trained checkpoints; times are on one NVIDIA A100-SXM4-40GB. Channel =ϕenc+ϕdec= _enc+ _dec; the proprioceptive encoder (4,6724,672 parameters per arm) is counted with the action head. Percentages are relative to the frozen backbone (9393M Octo-Base, ∼3.3 3.3B π0.5 _0.5). Component Value Backbone feature (i f_i) Octo-Base readout_action token, d=768d=768, frozen Proprio encoder (ϕprop _prop) Dense(7→64)(7→ 64)–ReLU–Dense(64→64)(64→ 64); dropout p=0.1p=0.1 in training Self representation (i h_i) ℝ832R^832 (wrist-only); ℝ1600R^1600 with the overhead feature π0.5 _0.5 state input Own 77-d state as prompt tokens (CoLA and LoRA alike); CoLA also via ϕprop _prop Message encoder (ϕenc _enc) Dense(832→96)(832→ 96)–ReLU–Dense(96→64)(96→ 64), dm=64d_m=64 (input 16001600 with overhead) Message decoder (ϕdec _dec) Dense(64→96)(64→ 96)–ReLU–Dense(96→768)(96→ 768), one per ordered pair Joint denoiser (ϕunet _unet) 1D U-Net, channels (128,256)(128,256), kernel 55, GroupNorm(8)(8), Mish; FiLM from Dense(256)(256)–Mish–Dense(256)(256) on [emb64(τ)∥i][emb_64(τ)\,\|\, z_i] Gripper head (ϕgrip _grip) Dense(i→128)( z_i→ 128)–Mish–Dense(128→K)(128→ K), binary logits Diffusion and loss S=100S=100, cosine schedule, ϵε-prediction, 1010-step DDIM; ϵε-MSE + 0.25+\ 0.25\,BCE Trainable: 2-arm and tray, wrist 14.4714.47M (15.6%15.6\%), channel 0.330.33M (0.36%0.36\%) Trainable: 2-arm, overhead 15.2015.20M (16.3%16.3\%), channel 0.480.48M (0.52%0.52\%) Trainable: 3-arm, overhead 23.9323.93M (25.7%25.7\%), channel 0.960.96M (1.04%1.04\%) Trainable: tray, π0.5 _0.5 16.9316.93M (∼0.5% 0.5\%), channel 0.830.83M (∼0.03% 0.03\%) Optimisation Adam (β=(0.9,0.999)β=(0.9,0.999), ϵ=10−8ε=10^-8, no weight decay), clip 1.01.0; cosine LR 10−4→10−510^-4→ 10^-5 over 150150 epochs; batch 128128 Model selection Best validation loss; early stopping patience 2525 Episodes (train/val/test) 2-arm 120/15/15120/15/15; tray 120/15/15120/15/15; 3-arm 104/13/13104/13/13 Compute Feature caching ∼5 5–1212 min per dataset; training ∼35 35 min (2-arm), ∼45 45 min (3-arm), ∼11 11 min (π0.5 _0.5) B.2 Language Instructions Table 4: Language instructions. The box is red in every task, and in the tray task only its marker changes colour, so no instruction reveals the hidden information. The role-specific instructions give the baselines more task information than CoLA; on the tray task, the π0.5 _0.5 baseline’s Arm B is also told that a tray exists. Model Arm Instruction CoLA (Octo-Base and π0.5 _0.5), all tasks all coordinate with partner Octo baselines, decentralised A pick up the red box and hand it to the other arm B, C take the red box from the other arm Octo baselines, centralised – pick up the red box with one arm and hand it to the other π0.5 _0.5 LoRA baseline (tray) A hand the box to the other arm B take the box from the other arm and put it in a tray B.3 Trainable Capacity and Training Cost Table 5: Trainable capacity and training cost by task and method. Every method within a task uses the same optimiser steps at batch size 128128, which matches the optimisation budget but not compute or capacity: CoLA trains on cached backbone features, so it does a fraction of the work per step. Trainable parameters are the modules updated by gradient descent, summed over the policies a method deploys (two for the 2-arm tasks, three for the 3-arm task); the Octo baselines’ action head is U-Net at horizon 1212 (68.568.5M), and the centralised variants differ only in output width (7→147→ 14), adding <0.1<0.1M. GPU-hours are job wall-clock on one NVIDIA A100-SXM4-40GB, summed over the policies a method deploys; CoLA trains all arms in a single model. They exclude CoLA’s one-time feature extraction (55–1212 min per Octo-Base dataset, amortised over every CoLA run and ablation on that dataset). The centralised policy encodes both arms’ cameras in every forward pass, which is why on the tray task it costs more than the two decentralised policies combined. Task Method Steps Trainable GPU-h Occluded tray handover (wrist only), 2 arms CoLA (with message) 20,25020,250 14.514.5M 0.60.6 CoLA (message zeroed) 20,25020,250 14.514.5M 0.60.6 Decentralised, head-only 20,25020,250 137.0137.0M 6.06.0 Centralised, head-only 20,25020,250 68.568.5M 6.46.4 Decentralised, full fine-tune 20,25020,250 323.0323.0M 6.16.1 Centralised, full fine-tune 20,25020,250 161.5161.5M 6.76.7 2-arm handover (overhead) CoLA (with message) 12,60012,600 15.215.2M 0.60.6 CoLA (message zeroed) 12,60012,600 15.215.2M 0.60.6 Decentralised, head-only 12,60012,600 137.0137.0M 3.73.7 Centralised, head-only 12,60012,600 68.568.5M 2.12.1 Decentralised, full fine-tune 12,60012,600 323.0323.0M 3.83.8 Centralised, full fine-tune 12,60012,600 161.5161.5M 2.12.1 3-arm handover chain (overhead) CoLA (with message) 26,85026,850 23.923.9M 0.80.8 CoLA (message zeroed) 26,85026,850 23.923.9M 0.80.8 Decentralised, head-only 26,85026,850 205.5205.5M 11.8†11.8 Decentralised, full fine-tune 26,85026,850 484.5484.5M 12.0†12.0 Occluded tray handover, π0.5 _0.5 backbone CoLA (with message) 20,25020,250 16.916.9M 0.20.2 Decentralised, LoRA fine-tune 20,25020,250 934.0934.0M 48.848.8 † Per-arm job wall-clock: head-only 3.953.95, 3.943.94 and 3.943.94 h; full fine-tune 4.024.02, 4.014.01 and 4.014.01 h. Appendix C Task Suite C.1 Qualitative Trajectory Comparison Figure 3: Qualitative Trajectory Comparison on Handover. Top: CoLA enables Arm A to maintain grasp stability until Arm B firmly closes its gripper. Bottom: Uncoordinated baseline desynchronises, causing premature release and dropped objects. Appendix D Full Experiment Results D.1 Two-Arm Handover Results Table 6: 2-Arm handover (overhead camera), per evaluation seed. Success rate (%) per 200-episode block; the seed sets the randomised start state. Octo-Base baselines use their wrist camera and the overhead camera. Method 0 1000 2000 3000 4000 Mean ± std CoLA (with message) 88.0 91.0 91.5 92.5 90.0 90.6 ± 1.7 CoLA (message zeroed) 90.5 85.0 89.5 – – 88.3 ± 2.9 Decentralised, head-only 66.5 66.0 69.0 67.5 72.5 68.3 ± 2.6 Centralised, head-only 62.5 62.0 68.5 63.5 73.0 65.9 ± 4.7 Centralised, full fine-tune 91.0 91.0 94.5 94.0 90.0 92.1 ± 2.0 Decentralised, full fine-tune 78.0 85.5 81.0 82.5 85.0 82.4 ± 3.1 Table 7: 2-Arm handover (wrist cameras only), per evaluation seed. Success rate (%) per 200-episode block at evaluation-seed offsets 0, 1000 and 2000. Method 0 1000 2000 Mean ± std CoLA (with message) 75.5 77.0 72.0 74.8 ± 2.6 CoLA (message zeroed) 61.5 59.5 72.0 64.3 ± 6.7 Decentralised, head-only 42.0 43.5 46.0 43.8 ± 2.0 Centralised, head-only 19.5 23.5 20.0 21.0 ± 2.2 Centralised, full fine-tune 81.5 81.5 77.0 80.0 ± 2.6 Decentralised, full fine-tune 71.0 69.5 67.5 69.3 ± 1.8 D.2 Three-Arm Handover Chain Results Table 8: 3-Arm handover chain (overhead camera), per evaluation seed. Success = both transfers completed and arm C holds the box for 20 steps; 200-episode blocks. CoLA rows have one block each; paired on the same 200 start states, message on vs. zeroed is 127 vs. 120 successes (exact McNemar p=0.50p=0.50). Method 0 1000 2000 3000 4000 Mean ± std CoLA (with message) 63.5 – – – – 63.5† CoLA (message zeroed) 60.0 – – – – 60.0† Decentralised, head-only 20.0 16.5 21.0 21.0 24.0 20.5 ± 2.7 Decentralised, full fine-tune 36.0 24.0 33.0 32.0 33.0 31.6 ± 4.5 Table 9: 3-Arm chain, per-stage rates (%), pooled over all evaluated blocks: each conditional is computed as the total over the total across blocks, not as a mean of per-block rates. A→ : first transfer. B→ ∣ → : second transfer among episodes that completed the first. Held∣ → : arm C holds the box for the required 20 steps, among episodes that completed the second transfer. Success is the product of the three stages, e.g. 42.9×51.3×93.2=20.542.9× 51.3× 93.2=20.5 for the head-only baseline. Method Blocks n Success A→ B→ ∣ → Held∣ → CoLA (with message) 1 200 63.5 79.5 79.9 100.0 CoLA (message zeroed) 1 200 60.0 75.0 80.0 100.0 Decentralised, head-only 5 1000 20.5 42.9 51.3 93.2 Decentralised, full fine-tune 5 1000 31.6 50.1 64.1 98.4 D.3 Occluded Tray-Handover Results Table 10: Occluded tray task, seed-0 block. All methods on the same 150 episodes (seed block 0, 50 per marker colour). Success = correct tray over all episodes, with 95% Wilson interval. Handover and Reached tray are rates over all episodes; Correct∣ and Correct∣ condition on those stages (chance on Correct∣ is 33.3%). Method n Success Handover Reached tray Correct∣ Correct∣ CoLA (with message) 150 76.0 [68.6, 82.1] 92.7 88.7 82.0 85.7 CoLA (message zeroed) 150 1.3 [0.4, 4.7] 8.0 6.0 16.7 22.2 Decentralised, head-only 150 22.7 [16.7, 30.0] 88.7 86.0 25.6 26.4 Centralised, head-only 150 16.0 [11.0, 22.7] 68.0 43.3 23.5 36.9 Centralised, full fine-tune 150 10.7 [6.7, 16.6] 79.3 10.7 13.4 100.0 Decentralised, full fine-tune 150 12.0 [7.7, 18.2] 94.7 36.7 12.7 32.7 CoLA on frozen π0.5 _0.5 150 37.3 [30.0, 45.3] 53.3 52.0 70.0 71.8 Decentralised, LoRA π0.5 _0.5 150 34.7 [27.5, 42.6] 100.0 100.0 34.7 34.7 Table 11: CoLA ablations on the occluded tray task. Message width: the same 150 episodes (seeds 0, 50 per marker colour) for every width. Lighting: out-of-distribution scene lighting, 30 episodes each. Ablation Setting n Correct Correct∣ Message width dmd_m 16 150 73.3 79.7 Message width dmd_m 32 150 42.0 47.4 Message width dmd_m 64 (default) 150 76.0 82.0 Message width dmd_m 128 150 73.3 76.9 Lighting (OOD) bright 30 3.3 50.0 Lighting (OOD) dark 30 10.0 15.8 Lighting (OOD) dim 30 13.3 19.0 D.4 Linear Probes of Channel Content Table 12: Linear probes on CoLA messages. Tray: multinomial logistic regression to marker colour, split by episode (chance 33.3%); frame = per-timestep accuracy, episode = majority vote. Handover: linear regression to box position. “Zeroed” probes the encoder of the model trained with the channel severed. Task / protocol Representation Dim Frame acc. Episode acc. Tray, 5-fold CV (150 ep.) self_a 832 100.0 100.0 Tray, 5-fold CV (150 ep.) msg_a 64 44.2 73.3 Tray, 5-fold CV (150 ep.) msg_a (zeroed model) 64 39.7 33.3 Handover (wrist only) R2R^2 (y / z) RMSE cm (y / z) Box position self_a 832 1.000 / 1.000 0.04 / 0.08 Box position msg_a 64 1.000 / 0.985 0.12 / 0.45 Box position msg_a (zeroed model) 64 0.999 / 0.982 0.13 / 0.50 Pose probe. On the wrist-only 2-arm handover, Arm A’s presentation pose (y,z)(y,z) is decodable from self (R2=1.000/ 1.000R^2=1.000\,/\,1.000), message (1.000/ 0.9851.000\,/\,0.985), and control (0.999/ 0.9820.999\,/\,0.982) alike (fig. 4, right; table 12). Because Arm A rigidly grasps the object, pose is near-linear in its proprioception and survives even a random projection, so the probe cannot attribute pose to training through the channel. It does, however, show that pose is reliably present in the message, and whether it matters depends on what Arm B can already see. With the overhead view, B observes the handoff directly and zeroing the message leaves success unchanged (90.6%90.6\% vs. 88.3%88.3\%). With only a wrist view, the message becomes B’s main source of A’s pose, and zeroing it drops success from 74.8%74.8\% to 64.3%64.3\%. Figure 4: Linear probes of channel content. Left: on the occluded tray-handover task, the message decodes box colour well above chance while an untrained control does not, indicating channel-specific information transfer. Right: on the wrist-only handover, the message and an untrained control both decode Arm A’s presentation pose near ceiling, so pose is linearly recoverable from the raw features regardless of training, and this probe cannot isolate the channel’s contribution. D.5 Zero-Shot Pretrained Octo Table 13: Zero-shot pretrained Octo (no fine-tuning; Bridge action statistics; overhead camera; 200 episodes, 350 control steps). Full task: pick up and hand over. HO: episodes start with arm A holding the box. Backbone Task Success Transfer A touches box B touches box Octo-Small Full task 0.0 0.0 0.0 0.0 Octo-Small HO 0.0 0.0 29.0 0.0 Octo-Base Full task 0.0 0.0 2.5 1.0 Octo-Base HO 0.0 0.0 11.0 0.0 D.6 Validation Gripper Accuracy Table 14: Validation gripper accuracy (%) per arm, taken at each run’s best-validation epoch, for the CoLA runs behind table 1. —: arm not present in the task. Task Arm A Arm B Arm C 2-Arm handover (wrist only) 98.8 98.8 — 2-Arm handover (overhead) 98.9 98.7 — Occluded tray handover 100.0 100.0 — 3-Arm handover chain 99.7 99.5 99.8 Appendix E Additional Training Dynamics fig. 5 shows validation loss for every training run behind table 1. All runs use the matched budget of 150150 epochs at batch size 128128; the Octo baselines’ optimiser steps are converted to epochs by dividing by that task’s steps per epoch (8484, 135135 and 179179 for the two-arm, tray and three-arm tasks). No CoLA run triggered early stopping, and the best validation epoch ranged from 131131 to 149149 across all CoLA runs, so every run consumed its 150150-epoch budget in full. Two patterns are worth noting. First, the gap between CoLA with messages and with messages zeroed is present on every task but varies several-fold in relative terms, and its ordering matches the behavioural results: final validation loss is 0.00090.0009 vs. 0.00390.0039 on the occluded tray task (4.3×4.3×), against 0.02430.0243 vs. 0.02830.0283 (1.2×1.2×) on the two-arm overhead task, where zeroing messages costs nothing behaviourally. Second, loss does not rank methods the way success rate does: the centralised policies have the highest validation loss on every task, since fitting a 1414-D joint action is harder than fitting a 77-D one, yet centralised full fine-tuning attains the best success rate on the two-arm overhead task. This is why we report success rates rather than losses throughout. Figure 5: Validation loss for all training runs in table 1, one panel per coordination regime, ordered as in the main table. Octo steps are converted to epochs so both families share the matched 150150-epoch budget. Decentralised rows train one policy per arm; the line is their mean and the shading their range. Appendix F Failure Mode Taxonomy Systematic inspection of failed trials identifies three orthogonal failure modes: F.1 Upstream Grasp Failures The initial grasping agent (Arm A) fails to secure the object due to physics edge cases during spawning. This failure is unimanual and independent of communication. F.2 Temporal Desynchronisation Failures Observed chiefly in the two-arm handover. Arm A ungrasps before Arm B secures contact, or Arm B closes prematurely. F.3 Post-Transfer Release Failures Specific to the three-arm chain, where success additionally requires Arm C to hold the object for 2020 steps after the second transfer. Arm C completes that transfer but loses the object during the hold in 1515 of the 220220 completed second transfers for the decentralised head-only baseline and 55 of 321321 for decentralised full fine-tuning, and in none of CoLA’s 127127. This mode is minor next to the transfers themselves: most failed three-arm episodes end at A→ or B→ rather than during the hold (table 9).