Paper deep dive
CTS-MoE: Implicit Terrain Adaptation via Mixture-of-Experts for Perceptive Locomotion
Francisco Affonso, Matheus P. Angarola, Ana Luiza Mineiro, Aditya Potnis, Marcelo Becker, Girish Chowdhary
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 98%
Last extracted: 6/21/2026, 4:17:04 AM
Summary
CTS-MoE is a perceptive legged locomotion framework designed for discontinuous terrain using a multi-task reinforcement learning approach. It combines a dense Mixture-of-Experts (MoE) actor for behavior composition with a sparse multi-critic to prevent value interference. The framework utilizes a concurrent teacher-student setup for asymmetric distillation, allowing the agent to adapt to terrain implicitly via perception (depth data) without requiring explicit terrain labels or high-level selectors at deployment. Experiments on a Unitree Go1 robot demonstrate that CTS-MoE achieves lower tracking error and higher success rates compared to monolithic and hierarchical baselines across seen and unseen terrains.
Entities (6)
Relation Signals (4)
CTS-MoE → deployedon → Unitree Go1
confidence 100% · Experiments on a Unitree Go1 in simulation and on hardware
CTS-MoE → implements → Multi-Task Reinforcement Learning
confidence 100% · We propose CTS-MoE... Cast as multi-task reinforcement learning
CTS-MoE → trainedin → IsaacLab
confidence 100% · leverage the IsaacLab [17] simulation environment
CTS-MoE → uses → Mixture-of-Experts
confidence 100% · We propose CTS-MoE, which combines a dense mixture-of-experts actor with perception-based gating
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Perceptive legged locomotion over discontinuous terrain (e.g., stairs, gaps, and obstacles) requires adaptive behavior, as a single conservative gait cannot produce the anticipatory maneuvers needed for abrupt topology changes. Cast as multi-task reinforcement learning, this problem introduces a tension between sharing and separation. Tasks use a common locomotion base but have conflicting rewards, so a policy must share behavior while avoiding value interference. Prior work addresses only one side, with monolithic policies sacrificing specialization and hierarchical sub-policies sacrificing generalization across transitions and unseen terrain. We propose CTS-MoE, which combines a dense mixture-of-experts actor with perception-based gating to compose shared behaviors and a multi-critic with task-specific value heads to prevent interference. The model is trained end-to-end in a single-stage concurrent teacher-student setup that handles partial observability and avoids sequential distillation, with task labels used only during training. At deployment, routing depends solely on perception, allowing terrain adaptation without a high-level selector or terrain classifier. Experiments on a Unitree Go1 in simulation and on hardware across seen and unseen terrains show task-aware specialization, with lower tracking error and higher success rates than monolithic baselines. Project Website: this https URL .
Tags
Links
- Source: https://arxiv.org/abs/2606.19633v1
- Canonical: https://arxiv.org/abs/2606.19633v1
Trouble viewing inline? Open PDF directly →
Full Text
56,898 characters extracted from source content.
Expand or collapse full text
CTS-MoE: Implicit Terrain Adaptation via Mixture-of-Experts for Perceptive Locomotion Francisco Affonso1, Matheus P. Angarola2, Ana Luiza Mineiro2, Aditya Potnis1, Marcelo Becker2, Girish Chowdhary1 1 University of Illinois Urbana-Champaign 2 University of São Paulo Abstract Perceptive legged locomotion over discontinuous terrain (e.g., stairs, gaps, and obstacles) requires adaptive behavior, as a single conservative gait cannot produce the anticipatory maneuvers needed for abrupt topology changes. Cast as multi-task reinforcement learning, this problem introduces a tension between sharing and separation. Tasks use a common locomotion base but have conflicting rewards, so a policy must share behavior while avoiding value interference. Prior work addresses only one side, with monolithic policies sacrificing specialization and hierarchical sub-policies sacrificing generalization across transitions and unseen terrain. We propose CTS-MoE, which combines a dense mixture-of-experts actor with perception-based gating to compose shared behaviors and a multi-critic with task-specific value heads to prevent interference. The model is trained end-to-end in a single-stage concurrent teacher–student setup that handles partial observability and avoids sequential distillation, with task labels used only during training. At deployment, routing depends solely on perception, allowing terrain adaptation without a high-level selector or terrain classifier. Experiments on a Unitree Go1 in simulation and on hardware across seen and unseen terrains show task-aware specialization, with lower tracking error and higher success rates than monolithic baselines. Project Website: https://cts-moe.github.io/ †footnotetext: Correspondence to faffonso@illinois.edu Figure 1: Real-world deployment of CTS-MoE on a Unitree Go1 across diverse outdoor terrains, including stairs, climbing obstacles, slopes, and uneven ground. Trained entirely in simulation in a single-stage multi-task formulation, the framework composes specialized expert behaviors directly from perception, enabling implicit terrain adaptation without explicit terrain labels or a high-level selector at deployment. Keywords: Legged Locomotion, Multi-Task Reinforcement Learning 1 Introduction Reinforcement learning (RL) provides a powerful paradigm for control, modeling sequential decision-making as trial-and-error interactions in which agents learn from reward signals to optimize long-term goals [25]. Within this field, RL has been widely adopted as a robust framework for legged locomotion, enabling agents to implicitly incorporate complex physical phenomena directly in the policy [13], and to reason about terrain geometry through perceptual inputs [16]. Yet locomotion over discontinuous terrain, such as stairs, gaps, and obstacles, remains an open challenge, as it requires a single agent to adapt its behavior to distinct anticipatory actions in each case. Standard approaches rely on a single reward signal to train locomotion policies across diverse scenarios, focusing on representation learning to enhance generalization. For instance, asymmetric teacher–student frameworks use a privileged teacher encoder—with access to ground-truth information such as terrain elevation, contact forces, and friction—to guide a deployable student operating on onboard sensory inputs [27, 31]. Other works incorporate vision-based [1, 7] and map-based [9] methods to enrich terrain understanding. However, these often yield overly conservative behaviors and fail to produce specialized maneuvers across varied terrains [2]. We identify the root cause as the need for task-specific rewards, rather than improvements in representation alone. Although objectives like velocity tracking remain consistent across tasks, others require tailored tuning; stair climbing, for instance, demands more relaxed orientation constraints than flat-ground walking. In this context, formulating the problem within a multi-task reinforcement learning (MTRL) framework provides a promising direction, but it also introduces additional training challenges [12, 6]. First, the presence of multiple reward functions with differing weights leads to gradient conflicts in value function learning, causing dominant tasks to bias or collapse the value estimates [18]. Second, using a single monolithic policy across diverse tasks without explicit contextual grounding makes it difficult for the controller to distinguish which behavior is appropriate in each situation. These issues become especially pronounced on discontinuous terrain, where the controller must rapidly adjust its behavior in response to sudden changes in surface topology [11]. These two challenges pull in opposite directions; value interference must be contained, while overlapping behaviors should be shared. To address the lack of versatile behaviors, prior work explores hierarchical frameworks that decompose locomotion into specialized sub-policies governed by a high-level selector, mitigating gradient issues during decomposed training [28, 10]. Although this modularity supports distinct behavioral styles, the disjointed training of sub-policies and selectors yields suboptimal transitions and a lack of end-to-end coherence, as the high-level logic stays unaware of physical limits. Mixture-of-Experts (MoE) architectures enable end-to-end learning, allowing a policy to combine specialized behaviors through learned routing. However, existing MoE methods have not been fully explored in the perceptive locomotion setting for discontinuous terrain, where implicit adaptation must arise directly from sensor data rather than explicit terrain labels. To overcome these limitations, we propose Concurrent Teacher-Student with Mixture of Experts (CTS-MoE), a perceptive locomotion framework that enables implicit terrain adaptation on discontinuous topologies. Our method combines shared locomotion objectives with task-specific constraints to encourage both generality and specialization. A dense MoE actor learns to compose specialized behaviors through soft routing, while a sparse multi-critic isolates task-specific value estimation to reduce interference. Together, these components allow the policy to adapt implicitly to terrain changes without requiring explicit terrain classification at deployment. The key contributions of this work are as follows: • We present a concurrent teacher-student framework that extends asymmetric distillation to the perceptive MTRL setting, jointly training the representation and expert policies end-to-end. • We resolve the sharing–separation tension in multi-reward RL asymmetrically, a dense MoE actor composes shared behaviors, while task-specific value heads prevent reward interference. • We demonstrate that perception-conditioned routing enables implicit policy adaptation, yielding task-aware specialization and generalizing to unseen terrains without explicit task labels. 2 Related Work Teacher-Student Frameworks. Asymmetric teacher-student frameworks address partial observability by distilling a privileged teacher encoder into a deployable student operating on onboard sensory inputs [13]; Wang et al. [27] extend this by training both encoders concurrently with the policy, eliminating distribution mismatch from sequential distillation. While effective for single-task locomotion, these frameworks typically train a monolithic policy without perceptual integration, limiting their ability to specialize across behaviors that require task-specific reward design. Hierarchical Policy Architectures. Extending RL-based locomotion to diverse terrains requires policies that can produce distinct, task-specific behaviors. As discussed in Wu et al. [28] and Yang et al. [30], different locomotion subtasks (e.g., walking, fall recovery, obstacle climbing) require distinct reward formulations to induce specialized behaviors. This motivates an MTRL formulation; however, it introduces inherent challenges, as each task may generate competing learning signals. Hoeller et al. [10] and Wu et al. [28] train separate per-terrain policies with a high-level selector to activate the appropriate sub-policy. Since the selector and experts are trained independently, misclassification can cause locomotion failure, and reliance on predefined terrain categories limits generalization to unseen environments. The root issue is the absence of end-to-end joint training, without which neither selector nor experts can adapt to each other’s representations. Zhuang et al. [33] instead distill the sub-policies into a single unified policy via DAgger [20], but performance stays bounded by the prior sub-policies. Mixture-of-Experts. To enable specialized sub-policies trained jointly, prior work has explored MoE architectures, which combine multiple sub-networks through a learned gating mechanism. GMT [5] demonstrates that dense MoE can capture diverse motion patterns in human motion retargeting, and similar ideas have been applied to blind locomotion [11] and manipulation [8]. Mysore et al. [18] further demonstrate that sparse gated critics can mitigate value interference. However, their method assumes that the desired behavioral style is provided explicitly during action generation, which is difficult to satisfy when task identity must be inferred from partial and noisy observations. As a result, a gap remains in integrating these approaches for perceptive MTRL, where the goal is to achieve versatile locomotion from perceptual inputs using end-to-end MoE-based training. Method Perc. MoE Multi-Task J-Train CTS [27] ✓ Ego-Vision [1] ✓ Parkour Learning [33] ✓ ✓ VocaLoco [28] ✓ ✓ MoE-Loco [11] ✓ ✓ CTS-MoE (Ours) ✓ ✓ ✓ ✓ Table 1: Comparison of frameworks. Perc. denotes depth-based perception. J-Train indicates end-to-end optimization, avoiding sequential distillation or high-level selection. In contrast, we integrate sparse-gated critics within a concurrent teacher-student MTRL framework while jointly learning dense MoE routing and expert policies end-to-end. Because routing is conditioned on perception rather than task labels, adaptation emerges implicitly, yielding smoother expert transitions on discontinuous terrain and better generalization to unseen environments. 3 Problem Formulation We formulate our locomotion framework as a partially observable Markov decision process (POMDP), since the agent does not have direct access to the full state and must instead act from observations. Moreover, we frame the problem as a multi-task learning setting grounded in the bounded similarity between locomotion behaviors; although the tasks share common goals and regularization terms, they differ in style, requiring careful and task-specific design to achieve specialization [32]. To approximate full observability, we encode the observation history under supervision from privileged state information. This yields a standard RL setting in which a policy π(at∣zt)π(a_t z_t)***For simplicity, we assume the encoded observation forms a valid Markovian state; while not true in general, this is a common practical simplification [19]. selects actions at∈a_t from latent representations zt∈z_t , computed from observations ot∈o_t . The interaction between the agent and the environment is represented by the transition distribution p(ot+1∣ot,at)p(o_t+1 o_t,a_t). We define a reward signal consistent with our multi-task formulation as r(ct)(ot,at)r^(c_t)(o_t,a_t), where ctc_t denotes the task embedding (provided as a task ID during training), allowing task-specific rewards to be propagated at each step. The goal of RL under this formulation is to maximize the discounted return by learning a policy that optimizes: J(π)=τ∼π[∑t=0∞γtr(ct)(ot,at)],J(π)=E_τ π [ _t=0^∞γ^tr^(c_t)(o_t,a_t) ], (1) where τ denotes an on-policy trajectory and γ∈[0,1)γ∈[0,1) is the discount factor. Observation and Action Space. The observation oto_t includes only deployable information, constructed by concatenating a history of proprioceptive states (otpo^p_t) over H timesteps with the current depth data from the camera. Additionally, during training in simulation, we assume access to an augmented state sts_t that contains privileged information, including the ground-truth heightmap hth_t, the task identifier ctc_t, and environmental parameters ete_t (e.g., terrain properties). Table 4 provides a detailed description of both the observation space and the privileged information. The action space ⊂ℝ12A ^12 consists of target joint positions, which are executed via a low-level PD controller. 4 Method We present CTS-MoE, a framework to enable versatile behaviors within an MTRL formulation conditioned on perceptive information, as illustrated in Fig. 2. Given the partial observability of our robotic setup, we employ distillation through a concurrent teacher-student framework to address representation learning challenges. Building on this, we extend standard CTS [27] and PPO [23] to the multi-task setting by incorporating a MoE into both the value function and the policy, enabling complementary forms of specialization. Concurrent Teacher-Student Framework. To mitigate partial observability, we train a teacher encoder as an oracle that leverages privileged observations (sts_t), following CTS, while extending the framework to support MTRL and perceptive inputs. A shared policy operates on the concatenated latent representations from the teacher and student encoders, defined as ztt=ψθt(st)z^t_t=ψ^t_ _t(s_t) and zts=ψθs(ot)z^s_t=ψ^s_ _s(o_t), parameterized by θt _t and θs _s, respectively. This design allows the teacher to guide learning during training while preserving deployability through the student encoder. At each policy update, only the teacher and actor-critic parameters are optimized via reinforcement learning. Subsequently, a supervised distillation step aligns the student encoder with the teacher’s latent representations. This enables the policy to progressively adapt to the student’s representation, while restricting the student to gradients from the distillation objective, thereby avoiding interference with policy optimization. This concurrent training scheme ensures the actor is conditioned on both encoders throughout learning, eliminating the need for an additional training phase required in sequential methods, which are prone to distribution mismatch when the student visits states outside the teacher’s training distribution [20]. The complete procedure is detailed in Appendix A. Figure 2: Overview of the proposed CTS‑MoE framework. The teacher and student are trained jointly in an asymmetric setup; the teacher uses privileged information while the student learns from deployable observations via distillation. On the policy side, a dense MoE composes actors through a learned router, while a separate multi-critic provides task-conditioned values to stabilize PPO. Multi-Reward. We formulate locomotion as an MTRL problem, as a single reward function cannot capture the diverse task-specific constraints and stylistic requirements. Our reward design, summarized in Table 7 and inspired by prior work [28, 22, 29], consists of task-specific objectives, stylistic regularization, and collision penalties, combined as a weighted sum. To ensure consistency across tasks, we retain shared components and weights for core locomotion and regularization terms, while introducing task-specific objectives where necessary. For tasks with continuous dynamics (e.g., walking), we employ standard velocity tracking. In contrast, discrete obstacles (e.g., obstacle climbing, gap crossing) require brief leaps or pauses that conflict with a fixed velocity target and would penalize the agent for the very behavior needed to traverse the discontinuity; we therefore reward proximity to a waypoint projected from the command, paired with a velocity kernel that preserves motion dynamics without enforcing an unattainable speed profile. We consider six tasks during training: flat terrain locomotion, stair ascent and descent, obstacle climbing (up and down), and gap crossing. To assign task-appropriate rewards at each timestep, we introduce auxiliary task embeddings ctc_t, denoted in the reward as r(ct)r^(c_t), which establish a correspondence between tasks and their associated terrains, assuming a total of N tasks. Multi-Critic and Value Normalization. To prevent dominant tasks from biasing value estimates under the multi-reward formulation, we employ a sparse multi-critic architecture. Each task ctc_t is assigned a dedicated value head, denoted Vϕ(ct)V_φ^(c_t), such that each critic is updated only using trajectories generated under its corresponding reward function. This design isolates competing learning signals across tasks and mitigates interference in value estimation. However, since different tasks yield returns with substantially different magnitudes, we stabilize training with per-task return normalization following POPArt [26], and normalize advantages independently per task before the PPO objective, preventing tasks with higher reward variance from dominating the policy gradient. The full objective is provided in Appendix A. MoE Policy. For the actor, we adopt a dense MoE to capture task-specific dynamics while enabling knowledge sharing across related behaviors. Unlike the reward and value functions, the actor does not rely on task embeddings, as such information is unavailable at deployment time. Since many tasks exhibit structural similarities, the policy benefits from combining multiple experts rather than selecting a single one. A routing network assigns soft weights to each expert based on perceptual inputs, under the assumption that task relevance is closely tied to the surrounding geometry. Proprioceptive information is also incorporated to capture the robot’s state and motion dynamics. Both modalities are encoded into latent representations, which are used by the router and expert policies. πθ(at∣zt,otp)=∑e=1Ege(zt)⋅πe(zt,otp),∑e=1Ege(zt)=1,zt=[ztt;zts]=[ψθt(st);ψθs(ot)], _θ(a_t z_t,o^p_t)= _e=1^Eg_e(z_t)· _e(z_t,o^p_t), _e=1^Eg_e(z_t)=1, z_t=[z^t_t;z^s_t]=[ψ^t_ _t(s_t);ψ^s_ _s(o_t)], (2) where ge(zt)g_e(z_t) represents the routing weights produced by a soft gating network for the e-th expert, πe(zt,otp) _e(z_t,o^p_t) denotes the e-th expert policy among E experts, and θ parameterizes the full policy. During training, the actor is trained using both teacher and student latent representations; at deployment, only the student representation ztsz_t^s is used. Finally, to promote balanced utilization and specialization in the MoE, we include auxiliary losses (Eq. 3) on the router probabilities pi=softmax(l)ip_i=softmax(l)_i, encouraging input‑dependent routing (ℒent)(L_ent), balanced expert usage (ℒbal)(L_bal), and numerical stability of the router logits (ℒl)(L_l). ℒent=[−∑ipilog(pi)],ℒbal=E⋅∑i[pi]2,ℒl=[‖l‖2].L_ent=E\! [- _ip_i (p_i) ], _bal=E· _iE[p_i]^2, _l=E[\|l\|^2]. (3) 5 Experimental Results In this section, we describe the experimental setup and evaluate our method against relevant baselines in both simulation and real-world environments. We report performance across seen and unseen terrains using metrics such as success rate and velocity tracking error. Additionally, we analyze the temporal behavior of the experts, as well as their average usage across tasks. Experimental Setup. We use the Unitree Go1 robot as our primary platform and leverage the IsaacLab [17] simulation environment, which integrates RL development through the RSL-RL library [24]. Our proposed method is built on a modified version of these frameworks. We train all policies on the Delta cluster [4] using four NVIDIA A40 GPUs for 15,000 iterations (∼ 72h). Detailed specifications of the simulation setup and training configurations are provided in Appendix B. Baselines. We evaluate our method against three baselines adapted from prior work. To ensure a fair comparison, all baselines are trained using the same pipeline, with each method modified to incorporate the MTRL structure in the critic with multi-reward (see Appendix A.1 for ablation studies comparing variants without MTRL components). Furthermore, all approaches are implemented within the asymmetric learning framework of CTS, and the MoE actor uses E=6E=6. • CTS-Single [1]: Policy trained using the proposed method without an MoE actor. • Blind CTS-MoE [11]: Policy trained using the proposed method without student perception. • Blind CTS-Single [27]: Policy trained using the proposed method without perception in the student and without an MoE actor. Performance Comparison. We first evaluate the linear and angular velocity tracking errors of the proposed method and baselines on both terrains seen during training and unseen terrains, as summarized in Table 2. The tracking error for each episode is computed using Eq. 4, which incorporates the success rate into the metric. This avoids ambiguous interpretations where episodes that collapse early but briefly exhibit good tracking would otherwise yield misleadingly low error values. Additionally, Table 3 reports the success rate separately. η=2ηv+ηω3,ηk=(1−term)(1T∑t=1T‖kt−ktcmd‖)+term,for k∈v,ω,η= 2 _v+ _ω3, _k=(1-I_term) ( 1T _t=1^T\|k_t-k_t^cmd\| )+I_term, k∈\v,ω\, (4) where term∈0,1I_term∈\0,1\ indicates whether the episode terminates early. We evaluate each method using forward locomotion commands, as the perceptual inputs are predominantly forward-facing and provide limited information for backward traversal. Accordingly, in all experiments we set vx∈0.5,0.75,1.0v_x∈\0.5,0.75,1.0\ m/s and fix ωz=0 _z=0. We use six difficulty levels for each terrain, as reported in Appendix B. Reported metrics are averaged across all command conditions and 15 episodes. Table 2: Comparison of Combined Velocity Tracking Errors (η) Across Methods and Terrain Environments. Method Diff. F A D G CU CD SL∗ R∗ CTS-MoE (Ours) 0–1 0.09 ± 0.05 0.08 ± 0.02 0.07 ± 0.02 0.26 ± 0.10 0.26 ± 0.03 0.18 ± 0.04 0.07 ± 0.01 0.10 ± 0.01 2–3 0.10 ± 0.07 0.09 ± 0.02 0.08 ± 0.02 0.28 ± 0.04 0.26 ± 0.08 0.20 ± 0.07 0.08 ± 0.01 0.10 ± 0.01 4–5 0.15 ± 0.08 0.20 ± 0.25 0.11 ± 0.02 0.32 ± 0.16 0.57 ± 0.37 0.32 ± 0.15 0.07 ± 0.01 0.11 ± 0.02 CTS-Single [1] 0–1 0.10 ± 0.03 0.11 ± 0.03 0.08 ± 0.02 0.25 ± 0.24 0.27 ± 0.27 0.15 ± 0.05 0.11 ± 0.04 0.14 ± 0.04 2–3 0.12 ± 0.04 0.12 ± 0.02 0.09 ± 0.02 0.39 ± 0.36 0.25 ± 0.20 0.19 ± 0.03 0.16 ± 0.05 0.15 ± 0.05 4–5 0.13 ± 0.08 0.32 ± 0.34 0.13 ± 0.03 0.66 ± 0.39 0.65 ± 0.38 0.30 ± 0.25 0.16 ± 0.05 0.15 ± 0.04 Blind CTS-MoE [11] 0–1 0.12 ± 0.03 0.18 ± 0.04 0.11 ± 0.02 0.34 ± 0.34 0.23 ± 0.18 0.13 ± 0.03 0.12 ± 0.02 0.16 ± 0.04 2–3 0.14 ± 0.04 0.28 ± 0.27 0.13 ± 0.03 0.72 ± 0.40 0.37 ± 0.31 0.16 ± 0.10 0.12 ± 0.02 0.19 ± 0.04 4–5 0.14 ± 0.04 0.79 ± 0.36 0.21 ± 0.09 0.94 ± 0.21 0.91 ± 0.24 0.28 ± 0.26 0.11 ± 0.02 0.17 ± 0.04 Blind CTS-Single [27] 0–1 0.22 ± 0.04 0.17 ± 0.10 0.18 ± 0.04 0.30 ± 0.24 0.18 ± 0.10 0.20 ± 0.04 0.21 ± 0.04 0.23 ± 0.05 2–3 0.22 ± 0.07 0.16 ± 0.16 0.17 ± 0.03 0.78 ± 0.34 0.24 ± 0.19 0.20 ± 0.04 0.20 ± 0.04 0.18 ± 0.05 4–5 0.22 ± 0.05 0.50 ± 0.42 0.24 ± 0.05 0.99 ± 0.06 0.71 ± 0.38 0.29 ± 0.18 0.19 ± 0.05 0.13 ± 0.02 Note: The tracking metric evaluates the mean error defined as η=(2ηv+ηω)/3η=(2 _v+ _ω)/3, integrating early termination penalties via Eq. 4. Terrains: Flat (F), Ascend (A), Descend (D), Gaps (G), Climb Up (CU), Climb Down (CD), Slope (SL∗), Rough (R∗). Superscript ∗ indicates previously unseen environments. (a) Expert usage patterns across the different terrains. (b) Pairwise cosine similarity between expert actions. Figure 3: Expert specialization and action similarity of the MoE actor across terrain tasks. Table 3: Success Rate (%). Terrain Ours [1] [11] [27] F 100.0 100.0 100.0 99.7 A 97.0 93.3 71.3 85.0 D 100.0 100.0 99.7 100.0 G 98.3 69.0 40.0 40.3 CU 87.0 76.7 63.3 78.0 CD 99.7 96.3 96.0 98.0 SL∗ 100.0 100.0 100.0 100.0 R∗ 100.0 100.0 100.0 100.0 Terrains: Flat (F), Ascend (A), Descend (D), Gaps (G), Climb Up (CU), Climb Down (CD), Slope (SL∗), Rough (R∗). Superscript ∗ indicates previously unseen environments. Tables 2 and 3 show that CTS-MoE outperforms perceptive baselines, with the largest gains on terrains that demand anticipatory behavior. Success rate increases by 29.3 percentage points on gaps and 10.3 percentage points on climb-up compared to the prior perceptive baseline. In contrast, the MoE actor yields no improvement for the blind baseline, as the absence of perceptual input prevents distillation of meaningful gating from the teacher. These results indicate that the gains arise from the MTRL formulation enabling specialized behaviors, rather than representation learning alone. The benefits also extend to velocity tracking. Even on simpler terrains where all methods achieve near-saturated success rates, CTS-MoE attains lower tracking error and maintains this advantage across curriculum progression and on unseen terrains. This suggests that perception-conditioned expert composition not only improves success on highly specialized tasks but also reduces tracking error in scenarios that are less dependent on representation quality. Figure 4: Per-expert router weights tracked against distance along the long evaluation course. Specialization. To analyze the behavior of the router and experts, Fig. 3 presents expert usage patterns across terrains along with the cosine similarity between their output actions. These results indicate that the MoE actor achieves task-aware specialization, with individual experts capturing complementary control strategies adapted to terrain-specific demands. This supports the hypothesis that the framework learns implicit expert compositions for both seen and unseen terrains without relying on explicit task identification. Figure 4 shows the temporal evolution of expert usage along a track with both seen and unseen terrains. A pure-pursuit controller maintains alignment with the centerline. The patterns indicate that expert selection depends on both terrain and time-varying control mixing, enabling smooth transitions without discrete policy switching as in hierarchical approaches. Figure 5: Real-world deployment of CTS-MoE across outdoor terrains that mirror the training tasks. Real-World Evaluation. We evaluated the proposed framework in real-world outdoor environments, targeting scenarios that are similar to those in simulation but with natural variations (see Appendix C for hardware details). This allows us to evaluate both sim-to-real transfer and generalization, avoiding overfitting to a fixed, instrumented setup with identical measurements. To this end, we conducted the tasks depicted in Fig. 5 and measured the success rate over 10 trials. We trained the policies with domain randomization (see Appendix B) to support sim-to-real transfer. Performance is stronger on terrains where perceptual input and rapid reactivity are less critical, such as climb-down and slope. The remaining gap is mainly due to limitations in domain randomization and depth estimation quality. MTRL-based policies require more reactive and complex joint actions, which amplify sim-to-real actuator mismatch. Additionally, the stereo-based depth sensor degrades under outdoor lighting and for very close objects (e.g., << 0.2 m), further affecting performance. 6 Limitations We identify three key limitations: • Our policy requires greater joint reactivity than simpler walking controllers, which amplifies the sim-to-real mismatch. Despite domain randomization and a low-pass filter at inference, frequency analysis shows that the deployed motions do not preserve the harmonic gait patterns observed in simulation. Recent methods [3] may help address this issue. • Expert gating depends implicitly on perception, so noise in the depth camera and sensitivity to lighting conditions can lead to incorrect expert selection, for example, climbing when unnecessary, and to occasional jerky motions before recovery. Improved sensing could mitigate these effects. • Training the full pipeline, including simulated depth and expert networks, in a single-stage increases the sample inefficiency of on-policy reinforcement learning [14]. This limits reproducibility to setups that require high-quality GPUs to support large-scale training. 7 Conclusion In this paper, we present CTS-MoE, a perceptive MTRL framework for legged locomotion over discontinuous terrain. By combining a dense mixture-of-experts actor with a sparse multi-critic and training both end-to-end in a concurrent teacher-student setup, the policy separates competing task rewards while composing specialized behaviors directly from perception. This enables implicit terrain adaptation without explicit labels or a high-level selector at deployment, avoiding the brittle transitions of hierarchical controllers. We believe this recipe can extend to other domains, such as loco-manipulation, where tasks share structure and a single deployable policy must handle diverse and unstructured environments. Acknowledgments This work used the Delta system through allocation CIS251247 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program, which is supported by National Science Foundation grants #2138259, #2138286, #2138307, #2137603, and #2138296. This work was also supported in part by the São Paulo Research Foundation (FAPESP) under grants no. 2025/24481-5 and 2025/27983-1. References [1] A. Agarwal, A. Kumar, J. Malik, and D. Pathak (2023) Legged locomotion in challenging terrains using egocentric vision. In Proceedings of The 6th Conference on Robot Learning, p. 403–415. Cited by: §1, Table 1, 1st item, Table 2, Table 3. [2] M. P. Angarola, F. Affonso, and M. Becker (2025) Learning terrain-specialized policies for adaptive locomotion in challenging environments. In 2025 IEEE International Conference on Advanced Robotics (ICAR), p. 16–22. Cited by: §1. [3] F. Bjelonic, F. Tischhauser, and M. Hutter (2025) Towards bridging the gap: systematic sim-to-real transfer for diverse legged robots. arXiv preprint arXiv:2509.06342. Cited by: Appendix B, 1st item. [4] T. J. Boerner, S. Deems, T. R. Furlani, S. L. Knuth, and J. Towns (2023) Access: advancing innovation: nsf’s advanced cyberinfrastructure coordination ecosystem: services & support. In Practice and Experience in Advanced Research Computing 2023: Computing for the Common Good, p. 173–176. Cited by: §5. [5] Z. Chen, M. Ji, X. Cheng, X. Peng, X. B. Peng, and X. Wang (2025) Gmt: general motion tracking for humanoid whole-body control. arXiv preprint arXiv:2506.14770. Cited by: §2. [6] G. Cheng, L. Dong, W. Cai, and C. Sun (2023) Multi-task reinforcement learning with attention-based mixture of experts. IEEE Robotics and Automation Letters, p. 3812–3819. Cited by: §1. [7] X. Cheng, K. Shi, A. Agarwal, and D. Pathak (2024) Extreme parkour with legged robots. In 2024 IEEE International Conference on Robotics and Automation (ICRA), p. 11443–11450. Cited by: Appendix B, §1. [8] C. Hao, X. Zhai, Y. Liu, and H. Soh (2026) Abstracting robot manipulation skills via mixture-of-experts diffusion policies. In The Fourteenth International Conference on Learning Representations, Cited by: §2. [9] J. He, C. Zhang, F. Jenelten, R. Grandia, M. Bächer, and M. Hutter (2025) Attention-based map encoding for learning generalized legged locomotionAttention-based map encoding for learning generalized legged locomotion. Science Robotics, p. eadv3604. Cited by: Appendix B, §1. [10] D. Hoeller, N. Rudin, D. Sako, and M. Hutter (2024) ANYmal parkour: learning agile navigation for quadrupedal robots. Science Robotics, p. eadi7566. Cited by: §1, §2. [11] R. Huang, S. Zhu, Y. Du, and H. Zhao (2025) MoE-loco: mixture of experts for multitask locomotion. In 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), p. 14218–14225. Cited by: §1, Table 1, §2, 2nd item, Table 2, Table 3. [12] Y. Kong, G. Ma, Q. Zhao, H. Wang, L. Shen, X. Wang, and D. Tao (2025) Mastering massive multi-task reinforcement learning via mixture-of-expert decision transformer. In Proceedings of the 42nd International Conference on Machine Learning, p. 31379–31396. Cited by: §1. [13] A. Kumar, Z. Fu, D. Pathak, and J. Malik (2021) RMA: rapid motor adaptation for legged robots. In Proceedings of Robotics: Science and Systems, Cited by: §1, §2. [14] J. Levy, T. Westenbroek, and D. Fridovich-Keil (2025) Learning to walk from three minutes of real-world data with semi-structured dynamics models. In Proceedings of The 8th Conference on Robot Learning, p. 2061–2079. Cited by: 3rd item. [15] Z. Li, X. B. Peng, P. Abbeel, S. Levine, G. Berseth, and K. Sreenath (2025) Reinforcement learning for versatile, dynamic, and robust bipedal locomotion control. The International Journal of Robotics Research 44 (5), p. 840–888. Cited by: Appendix C. [16] G. B. Margolis, T. Chen, K. Paigwar, X. Fu, D. Kim, S. b. Kim, and P. Agrawal (2022) Learning to jump from pixels. In Proceedings of the 5th Conference on Robot Learning, p. 1025–1034. Cited by: §1. [17] M. Mittal, P. Roth, J. Tigue, A. Richard, O. Zhang, P. Du, A. Serrano-Munoz, X. Yao, R. Zurbrüg, N. Rudin, et al. (2025) Isaac lab: a gpu-accelerated simulation framework for multi-modal robot learning. arXiv preprint arXiv:2511.04831. Cited by: §5. [18] S. Mysore, G. Cheng, Y. Zhao, K. Saenko, and M. Wu (2022) Multi-critic actor learning: teaching RL policies to act with style. In International Conference on Learning Representations, Cited by: §1, §2. [19] Physical Intelligence, A. Amin, R. Aniceto, A. Balakrishna, K. Black, K. Conley, G. Connors, J. Darpinian, K. Dhabalia, J. DiCarlo, et al. (2025) π0.6∗π^*_0.6: A VLA that learns from experience. arXiv preprint arXiv:2511.14759. Cited by: footnote *. [20] S. Ross, G. Gordon, and D. Bagnell (2011) A reduction of imitation learning and structured prediction to no-regret online learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, p. 627–635. Cited by: §2, §4. [21] N. Rudin, D. Hoeller, M. Bjelonic, and M. Hutter (2022) Advanced skills by learning locomotion and local navigation end-to-end. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), p. 2497–2503. Cited by: Appendix B. [22] N. Rudin, D. Hoeller, P. Reist, and M. Hutter (2022) Learning to walk in minutes using massively parallel deep reinforcement learning. In Proceedings of the 5th Conference on Robot Learning, p. 91–100. Cited by: §4. [23] J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov (2017) Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347. Cited by: §4. [24] C. Schwarke, M. Mittal, N. Rudin, D. Hoeller, and M. Hutter (2025) RSL-rl: a learning library for robotics research. arXiv preprint arXiv:2509.10771. Cited by: §5. [25] R. S. Sutton and A. G. Barto (2018) Reinforcement learning: an introduction. 2nd edition, MIT Press, Cambridge, MA. Cited by: §1. [26] H. P. Van Hasselt, A. Guez, M. Hessel, V. Mnih, and D. Silver (2016) Learning values across many orders of magnitude. In Advances in Neural Information Processing Systems, Cited by: Appendix A, §4. [27] H. Wang, H. Luo, W. Zhang, and H. Chen (2024) CTS: concurrent teacher-student reinforcement learning for legged locomotion. IEEE Robotics and Automation Letters, p. 9191–9198. Cited by: Appendix A, §1, Table 1, §2, §4, 3rd item, Table 2, Table 3. [28] S. Wu, M. H. Danesh, S. Li, H. Yurchyk, A. Abyaneh, A. El Houssaini, D. Meger, and H. Lin (2025) VOCALoco: viability-optimized cost-aware adaptive locomotion. IEEE Robotics and Automation Letters, p. 1146–1153. Cited by: §1, Table 1, §2, §2, §4. [29] T. Wu, H. Guo, Y. Wang, J. Yang, X. Sui, J. Xie, X. Chen, Z. Liu, and X. Lan (2026) Toward reliable sim-to-real predictability for moe-based robust quadrupedal locomotion. In Proceedings of Robotics: Science and Systems, Cited by: Appendix B, §4. [30] C. Yang, K. Yuan, Q. Zhu, W. Yu, and Z. Li (2020) Multi-expert learning of adaptive legged locomotion. Science Robotics 5 (49), p. eabb2174. Cited by: §2. [31] Z. Zhang, J. Guo, C. Chen, J. Wang, C. Lin, Y. Lian, H. Xue, Z. Wang, M. Liu, J. Lyu, et al. (2025) Track any motions under any disturbances. arXiv preprint arXiv:2509.13833. Cited by: §1. [32] Z. Zhao, Z. Zhao, K. Xu, Y. Fu, J. Chai, Y. Zhu, and D. Zhao (2025) Learning and planning multi-agent tasks via an moe-based world model. In The Thirty-ninth Annual Conference on Neural Information Processing Systems, Cited by: §3. [33] Z. Zhuang, Z. Fu, J. Wang, C. G. Atkeson, S. Schwertfeger, C. Finn, and H. Zhao (2023) Robot parkour learning. In Proceedings of The 7th Conference on Robot Learning, p. 73–92. Cited by: Appendix C, Table 1, §2. Appendix A CTS-MoE Training Pipeline Name Symbol Dimension Proprioceptive (otp)(o^p_t) Lin. Vel. Command vtcmdv_t^cmd 2 Ang. Vel. Command ωtcmd _t^cmd 1 Base Ang. Vel. ωt _t 3 Proj. Gravity gtg_t 3 Joint Positions qtq_t 12 Joint Velocities q˙t q_t 12 Previous Actions at−1a_t-1 12 Perceptive Depth dtd_t 48 × 64 Heightmap* hth_t 187 Task Embedding Task ID* ctc_t 1 Environmental (et)(e_t) Base Velocity* vtv_t 3 Contact Forces* FcF_c 4 Joint Torques* τt _t 12 Joint Accel.* q¨t q_t 12 Robot Mass* m 1 Joint Stiffness* ksk_s 12 Joint Damping* kdk_d 12 Table 4: Observation and privileged state spaces. Elements marked with (*) are teacher-only information. Due to our method being built on CTS, this section provides background on how the framework works, as well as our modifications to adapt it to our MTRL formulation and perception-aware setup. We also present representation details describing how the networks are configured to realize the encoders introduced earlier, and detail the full optimization objectives, including the value loss formulation with per-task return normalization. Encoder Architectures and Representations. We train the teacher and student encoder networks using the information listed in Table 4. The teacher receives privileged information, combining proprioceptive, perceptive, and environmental components. These inputs are processed separately, as described in Eq. 5, so that each component can be encoded independently without over-compressing important signals. Specifically, the environmental privileged information ete_t is passed through an MLP, the perceptive heightmap component hth_t is processed by another MLP, and both representations are concatenated with the task embedding ctc_t through a linear layer followed by normalization. ztt=ψθt(st)=LN(MLP(et)⊕MLP(ht)⊕ct).z_t^t=ψ^t_ _t(s_t)=LN(MLP(e_t) (h_t) c_t). (5) To match this representation using deployable observations, we feed a network with the history of observations to estimate the proprioceptive information, and rely on a recurrent network (GRU) to predict the heightmap in a memory-based fashion, passing the depth data within the observation to guide the robot’s direction. The task identification is approximated by a mix of both proprioceptive and perceptive information through a linear transformation: zts=ψθs(ot)=LN(MLP(ot−H:tp)⊕GRU(CNN(dt)⊕otp)).z_t^s=ψ^s_ _s(o_t)=LN(MLP(o^p_t-H:t) (CNN(d_t) o^p_t)). (6) (a) Teacher Encoder. (b) Student Encoder. Figure 6: Comparison between the Teacher and Student encoder architectures. Optimization Objectives. To incorporate both the teacher-student framework and multi-reward approach, we modify the standard PPO. Following [27], since agents are divided into teacher and student groups, the Monte-Carlo approximation of PPO’s clipped objective is defined as: ℒPPO,t(θ,θt) _PPO,t(θ, _t) =1|t|∑τ∈t∑t=0Tmin(ρtA~t,clip(ρt,1−ϵ,1+ϵ)A~t), = 1|D_t| _τ _t _t=0^T ( _t^t A_t^t,\ clip( _t^t,1-ε,1+ε) A_t^t ), (7) ℒPPO,s(θ) _PPO,s(θ) =1|s|∑τ∈s∑t=0Tmin(ρtsA~ts,clip(ρts,1−ϵ,1+ϵ)A~ts), = 1|D_s| _τ _s _t=0^T ( _t^s A_t^s,\ clip( _t^s,1-ε,1+ε) A_t^s ), (8) where t,sD_t,s represents data collected from teacher/student groups, A~t,s A_t,s is the normalized advantage function for each group, T is the trajectory horizon, and ϵε is the clipping parameter. ρ represents the probability ratio function for the two groups: ρt(θ,θt) _t^t(θ, _t) =πθ(att|ott,ψθt(st))πθold(att|ott,ψθoldt(st)), = _θ(a_t^t|o_t^t,ψ^t_θ(s_t)) _ _old(a_t^t|o_t^t,ψ^t_ _old(s_t)), (9) ρts(θ) _t^s(θ) =πθ(ats|ots,ψθs(ot))πθold(ats|ots,ψθolds(ot)). = _θ(a_t^s|o_t^s,ψ^s_θ(o_t)) _ _old(a_t^s|o_t^s,ψ^s_ _old(o_t)). (10) Because tasks differ in reward magnitude, raw value targets and advantages can vary by orders of magnitude across tasks, destabilizing value learning. To address this, we apply per-task return normalization following POPArt [26]. For each task, an exponential moving average tracks the running return statistics, which are used to whiten both the targets and the critic predictions, we denote these normalized quantities as V~ϕ(k) V_φ^(k) and R~t(k) R_t^(k). The critic then minimizes a Huber loss in this whitened space, averaged over all N tasks: ℒvalue(ϕ)=1N∑k=1N(k)[ℋδ(V~ϕ(k)(zt),R~t(k))],L_value(φ)= 1N _k=1^NE_D^(k)\! [H_δ\! ( V_φ^(k)(z_t),\, R_t^(k) ) ], (11) where (k)D^(k) denotes transitions assigned to task k=ctk=c_t. When the moving average parameters are updated, the critic output layer is rescaled to keep the whitened prediction unchanged, following the standard POPArt update. On the actor side, advantages are likewise normalized per task before entering the PPO surrogate (Eq. 7 and 8), preventing tasks with larger reward variance from dominating the policy gradient. After each PPO update, we update the student encoder using trajectories from sD_s. We distill teacher latents into the student encoder via: ℒMSE(θs)=1|s|∑τ∈s∑t=0T‖ψθs(ot)−ψθt(st)‖22.L_MSE( _s)= 1|D_s| _τ _s _t=0^T \| _ _s^s(o_t)- _ _t^t(s_t) \|_2^2. (12) This prevents the student from learning the same RL signal as the teacher, avoiding policy training contamination. A.1 Ablation Study: Is Task-Specific Reward Design Necessary? To support the hypothesis that an MTRL formulation is necessary to achieve more diverse scenarios through task-specific rewards, rather than through improvements in representation alone, we evaluate an oracle agent with access to privileged information. We benchmark variants of this oracle across single- and multi-reward components, as well as single- and multi-critic components. To simplify the evaluation, we train to convergence on a restricted curriculum of terrains, using flat as a representative standard task and climb-up as a complex task that requires task-specific terms to enable greater flexibility (e.g., orientation) to achieve the desired behavior and explore the state space sufficiently to succeed. We then evaluate on this curriculum as well as on rough terrain, an out-of-distribution task used to assess robustness. Specifically, we consider four configurations to assess each component independently, where the shared-reward variants rely on a unified reward across all tasks; we use the flat reward, the standard reward commonly adopted in monolithic formulations. • Multi-Critic with Multi-Reward (MC+MR). • Multi-Critic with Shared-Reward (MC+SR). • Shared-Critic with Multi-Reward (SC+MR). • Shared-Critic with Shared-Reward (SC+SR). As shown in Table 5, we report success rates across tasks, along with the average value loss during training to evaluate the effect of the multi-critic design. The results indicate that multi-reward formulations are critical for discontinuous terrains such as climb-up, where they achieve significantly higher success rates. This supports our hypothesis that task-specific reward structures are necessary to capture distinct behavioral requirements. Table 5: Ablation of Multi-Critic vs. Shared Critic showing Success Rates (%) and training value loss. Method Flat (↑ ) Rough (↑ ) Climb Up (↑ ) Overall (↑ ) Value Loss (↓ ) (×10−3× 10^-3) MC+MR (Ours) 97.6 54.6 71.7 74.6 4.54 SC+MR 89.3 38.1 67.5 65.0 4.78 MC+SR 91.2 51.4 42.9 61.8 3.33 SC+SR 86.2 29.8 44.2 53.4 7.15 Additionally, the value loss analysis shows that the multi-critic consistently reduces the overall critic loss, indicating more accurate value estimation during PPO training. This effect is more pronounced in the shared-reward setting, where the absence of task-specific rewards forces a single reward function to be shared across tasks. In such cases, the multi-critic design provides a larger relative improvement, suggesting that it is particularly beneficial when task inference is more challenging; the same reward function induces different value functions across tasks due to the differing transition dynamics of their terrains. Finally, the lower absolute value loss observed in the MC+SR setting is attributed to the reduced curriculum level it reaches compared to multi-reward formulations, resulting in an overall simpler learning problem. Appendix B Simulation Setup This appendix provides additional details regarding our simulation setup to facilitate reproducibility. We include comprehensive specifications for the network architectures, training hyperparameters, terrain configurations, and reward formulations used in our experiments. Network Architectures. Table 6 details the architectural specifications and layer dimensions for all components in our framework, ranging from the input encoders to the policy and critic networks. Table 6: Network architecture. Component Input Output Hidden Layers Teacher Encoder ψθtψ^t_ _t Privileged MLP ete_t 32 [512, 256] Heightmap MLP hth_t 128 [512, 256] Latent Projection ct⊕MLP(et)⊕MLP(ht)c_t (e_t) (h_t) 32 Linear + LN Student Encoder ψθsψ^s_ _s Proprioception MLP ot−H:tpo^p_t-H:t 32 [512, 256] Depth CNN ϕφ dtd_t 128 Filters [16, 32, 64] Temporal GRU ϕ(dt)⊕otpφ(d_t) o^p_t 256 1 layer Latent Projection GRU(⋅)⊕MLP(ot−H:tp)GRU(·) (o^p_t-H:t) 32 Linear + LN Policy and Value Heads πθVϕ _θ V_φ Router MLP ztz_t E [512, 256] Actors (×E× E) zt⊕otpz_t o^p_t |||A| [512, 256, 128] Critics (×N× N) zt⊕otpz_t o^p_t 1 [512, 256, 128] Reward Formulation. The reward weights for each specific task are detailed in Table 7. Obstacle-related tasks are guided by point tracking, because for agile maneuvers such as climbing up, climbing down, and crossing gaps, utilizing projected goal points is essential because it grants the locomotion policy the freedom to dynamically modulate its velocity and gait, overcoming the rigid constraints of continuous velocity tracking [7, 21]. Table 7: MTRL Reward Weights by Terrain Task Reward Term Flat Ascend Descend Cross Gap Climb Up Climb Down Shared Task Components rvxyr_v_xy 2.0 1.5 1.5 – – – rωzr_ _z 1.0 0.75 0.75 – – – rwpr_wp – – – 1.0 1.0 1.0 Shared Regularization rvzr_v_z -1.5 -0.25 -1.5 -1.0 -0.25 -1.5 rωxyr_ _xy -0.05 -0.05 -0.05 -0.05 -0.05 -0.05 rqlimr_q_lim -2.0 -2.0 -2.0 -2.0 -2.0 -2.0 rq¨r_ q -2.5e-7 -2.5e-7 -2.5e-7 -2.5e-7 -2.5e-7 -2.5e-7 rτr_τ -1.0e-4 -1.0e-4 -1.0e-4 -1.0e-4 -1.0e-4 -1.0e-4 rPr_P -2e-5 -2e-5 -2e-5 -2e-5 -2e-5 -2e-5 ra˙r_ a -0.01 -0.01 -0.01 -0.01 -0.01 -0.01 ra¨r_ a -0.01 -0.01 -0.01 -0.01 -0.01 -0.01 Task-Specific Stylistic Terms rcr_c -0.5 -1.0 -1.0 -1.0 -1.0 -1.0 rqhipr_q_hip -0.05 -0.05 -0.05 -0.05 -0.05 -0.05 rfeetr_feet -0.05 -0.2 -0.1 -0.05 -0.2 -0.1 rorir_ori -1.0 – – – – – rhr_h -10.0 -10.0 -10.0 -10.0 -10.0 -10.0 rtermr_term -5.0 -5.0 -5.0 -20.0 -20.0 -20.0 Hyperparameters. We extend PPO with teacher-student groups using a distillation process after each update, and MoE auxiliary losses controlling load balancing, router entropy, and l-loss regularization. Table 8 summarizes the corresponding hyperparameters. Table 8: Summary of training hyperparameters. Category Parameter Value PPO Clip ratio 0.2 Learning rate 10−310^-3 Discount factor 0.99 GAE-lambda 0.95 Desired KL divergence 0.01 Max gradient norm 1.0 Learning epochs / Mini-batches 5 / 4 Value Loss Coefficient 1.0 Entropy Coefficient 0.005 Group Proportions of Envs Teacher 0.85 Student 0.15 MoE Auxiliary Balance Coefficient 2×10−32× 10^-3 Router Entropy Coefficient 5×10−45× 10^-4 Router l-Loss Coefficient 1×10−51× 10^-5 Distillation (Student) Behavior Loss 1.0 Task-Dependent Commands: Target commands (vx,vy,ωzv_x,v_y, _z) are uniformly sampled based on task. For velocity tracking tasks, broad ranges are applied (e.g., vx∈[−1.0,1.0]v_x∈[-1.0,1.0] m/s, vy∈[−0.5,0.5]v_y∈[-0.5,0.5] m/s, ωz∈[−0.8,0.8] _z∈[-0.8,0.8] rad/s). In contrast, point tracking terrains (e.g., obstacles, gaps) restrict sampling strictly to forward velocities (vx∈[0.0,1.0]v_x∈[0.0,1.0] m/s) to ensure alignment with waypoints. Terrain Generation and Curriculum. To ensure that the distribution of training conditions reflects the variety of surfaces expected at deployment, we construct a comprehensive elevation map composed of 42 sub-terrains arranged in a 7 × 6 grid. Each row corresponds to a specific terrain type with progressively increasing difficulty. Each terrain patch spans 8m × 8m and begins with a 2m × 2m initial platform. The generated terrains are illustrated in Fig. 7. Figure 7: Overview of the training sub-terrains organized by terrain type and escalating difficulty. Table 9: Terrain curriculum parameters. Task Curriculum Parameter Difficulty Range Seen Terrain Flat — — Descend Step Height [0.02,0.22][0.02,0.22] m Ascend Step Height [0.02,0.22][0.02,0.22] m Cross Gap Gap Thickness [0.05,0.50][0.05,0.50] m Obstacles Obstacle Height [0.05,0.50][0.05,0.50] m Unseen Terrain Rough Frequency / Height Scale [18,40][18,40] / [0.12,0.58][0.12,0.58] m Slope Slope Incline [0,30]∘[0,30] We guide the training process using a performance-based, probabilistic terrain curriculum. A success is registered if the robot reaches the target goal or traverses over half the terrain length, while failure occurs if it travels less than half of its expected distance. To mitigate catastrophic forgetting and maintain experience diversity, difficulty transitions are stochastic rather than strictly linear. Following a success, the environment has an 80% probability of advancing to the next difficulty level and a 20% chance of resetting to a random easier level. Conversely, failures result in a difficulty downgrade, which incorporates a 10% probability of a complete reset to the base terrain to facilitate policy recovery. Figure 8: Unseen terrains not included during training to assess generalization: rough, slope up, and slope down. Additionally, completely unseen terrains were created specifically for the experimental evaluation. As illustrated in Fig. 8 and described in Table 9, this testing set consists of a withheld subset of rough, slope up, and slope down terrains. Domain Randomization. To ensure robust hardware deployment, we first present simulation results using the simulator’s integrated actuators, which are efficient but poorly represent real actuators, hindering sim-to-real transfer. After convergence under this setup, we resume training with extensive domain randomization, as in [9], relying on an explicit delayed PD actuator model so the policy accounts for actuator dynamics, control latency, and non-ideal hardware. Parameters follow [29], given the similar terrain and platform. Despite these efforts, and as noted in the limitations, we found that high-agility tasks would benefit from further techniques such as [3]. Training metrics. Figure 9 reports the reward as a function of simulation steps, while Table 10 summarizes the maximum curriculum level reached by each method. The reward alone provides limited insight, as it quickly converges to a stable value and remains nearly constant throughout training. This behavior arises from the evolving curriculum, i.e., as task difficulty increases, similar reward values correspond to progressively more challenging scenarios. Examining the curriculum results reveals that policies with an MoE actor consistently reach higher difficulty levels, suggesting that the architecture provides implicit terrain adaptation. In contrast, under blind settings, the absence of perceptual inputs prevents effective distillation to the student, since the student lacks access to the perceptive information required to approximate the teacher’s behavior. The curriculum progression makes this gap more apparent. Finally, the teacher in blind settings attains a slightly higher curriculum level. We attribute this to differences in the training pipeline. Perceptive settings involve additional networks trained under sequential constraints (e.g., due to the GRU-based recurrent module), which introduces optimization challenges and slightly degrades oracle teacher performance. Nevertheless, this drawback is offset by improved student performance in the perceptive setting. Figure 9: Average reward and standard deviation over simulation steps across methods. Table 10: Maximum Curriculum Levels. CTS-MoE (Ours) CTS-Single Blind CTS-MoE Blind CTS-Single Terrain Teacher Student Teacher Student Teacher Student Teacher Student Climb Up 4.35 4.11 4.02 4.00 4.99 3.99 4.24 4.01 Climb Down 6.00 6.00 6.00 6.00 6.00 6.00 6.00 6.00 Flat 6.00 6.00 6.00 6.00 6.00 6.00 6.00 6.00 Descend 6.00 6.00 6.00 6.00 6.00 6.00 6.00 6.00 Ascend 5.91 5.60 5.10 5.04 5.85 5.17 5.92 5.07 Cross Gap 6.00 6.00 6.00 6.00 6.00 4.15 6.00 3.95 Average 5.71 5.62 5.52 5.51 5.81 5.22 5.69 5.17 Appendix C Hardware Setup Figure 10: Unitree Go1 quadruped robot equipped with an Intel RealSense D435i depth camera. Our experimental setup is based on the Unitree Go1 quadruped robot, featuring 12 degrees of freedom (see Fig. 10). Exteroceptive perception relies on a head-mounted Intel RealSense D435i depth camera. Due to the computational demands of the depth, CNN, and GRU processing, the full policy runs on an on-board Jetson Xavier NX at 50 Hz. The resulting target joint positions are sent directly to a low-level PD controller (Kp=28K_p=28, Kd=0.65K_d=0.65) operating at 250 Hz. The camera setup follows [33], using a forward-facing view tilted about 30∘30 downward to capture both nearby footholds and upcoming terrain. We additionally apply a second-order low-pass filter to the policy outputs, following [15], with a cutoff frequency of 12 Hz.