acceptodds
Under review as a conference paper at ICLR 2027

P-BFM: Perceptive Behavioral Foundation Models for Humanoid Control via Unsupervised Reinforcement Learning

Abstract

Behavioral foundation models (BFMs) provide humanoids with reusable repertoires of whole-body behaviors; in real-world humanoid control, the same behavioral intent may require fundamentally different motions under different terrain constraints. Covering such variation with terrain-matched demonstrations requires enumerating a combinatorial space of behaviors and terrain geometries. We introduce P-BFM, a perceptive behavioral foundation model that factorizes humanoid control into terrain-invariant intent and terrain-conditioned realization. A forward–backward representation maps motion references, target states, and reward functions into a shared intent space without terrain information, while a perceptive policy learns how to realize each intent under the surrounding terrain geometry. We further identify a complementary limitation of purely latent control: behavioral latents can preserve local motion patterns while losing precise global progress. We therefore introduce Explicit Root Instructions (ERI), which separates global transport (e.g., heading and planar velocity) from behavioral intent, while leaving local whole-body adaptation to the perceptive policy. This decomposition leads to a characteristic capability absent from conventional BFMs: one fixed intent induces different, terrain-appropriate behaviors, including changes in step height, stride, posture, foothold placement, and contacts, without terrain- specific demonstrations. Across challenging terrains, P-BFM improves motion fidelity, root-control precision, and perturbation recovery over strong baselines. A single frozen policy further supports motion tracking, goal reaching, reward inference, teleoperation, and terrain-aware recovery, and transfers zero-shot to a Unitree G1 using onboard depth sensing.

Primary area:
applications to robotics, autonomy, planning
Keywords:
behavioral foundation models, humanoid control, terrain adaptation, perceptive control, unsupervised reinforcement learning, forward-backward representations

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_JCBiOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces P-BFM, a perceptive behavioral foundation model for humanoid control that realizes shared behavioral intents across different terrains without terrain-matched demonstrations. Built on forward–backward representations, it maps motion references, target body states, and reward functions into a terrain-agnostic intent space. A depth-conditioned policy then adapts execution to the surrounding terrain. Explicit Root Instructions provide heading and planar-velocity targets to improve global motion precision. A single frozen policy supports motion tracking, goal reaching, reward inference, and teleoperation.

Critical strengths
  1. The central design is clear and offers an element of novelty. Separating behavioral intent from terrain-conditioned execution allows the policy to preserve the intended behavior while adjusting step height, footholds, posture, and contacts. This decomposition is well aligned with the requirements of humanoid control across terrains.

  2. Reducing reliance on terrain-matched demonstrations has practical value. The method uses flat-ground motions as behavioral prompts and motion priors, while learning terrain adaptation through autonomous interaction. This reduces the need to construct reference trajectories for individual motion–terrain combinations.

  3. ERI is well motivated and supported by direct ablations. The paper identifies that behavioral latents may preserve local motion patterns without ensuring accurate heading or forward progress. Explicit root instructions provide a simple complement, and the corresponding ablations support their contribution.

  4. The unified policy supports flexible task specification. Motion tracking, goal reaching, reward inference, and teleoperation share the same behavioral representation and execution policy without task-specific policy optimization.

Critical weaknesses

The ERI and perception ablations demonstrate the value of instructions and visual inputs, but do not directly establish whether excluding terrain information from the backward representation is important for the performance. A comparison in which only the backward representation’s access to terrain information changes would help isolate the benefit of terrain-agnostic intent encoding. This would strengthen the experimental support for decomposition of intent and terrain-conditioned execution.

Questions for authors

Can the contribution of terrain-agnostic intent encoding be evaluated directly? With the actor, perceptual inputs, ERI, and training settings held consistent, how would allowing the backward representation to receive terrain information affect motion fidelity and terrain adaptation?

2 comments
AuthorsOpenReview ↗

Rebuttal: Benefits of Separating Intent Encoding from Terrain Perception

Thank you for your thoughtful feedback and for recognizing our motivation and pipeline design. The response and additional results are also available on the anonymous rebuttal page.

Benefits of Separating Intent Encoding from Terrain Perception

1. Better tracking accuracy and robustness. As shown in the table below, adding depth to the backward encoder B reduces motion-tracking accuracy at inference time and makes the policy less robust. We see two possible reasons.

  1. During training, we use paired intent latents and motion capture clips to guide exploration. If the intent latent depends on depth, the motion prior becomes implicitly conditioned on flat ground. This can create a mismatch with the intents encountered on diverse terrains during training and weaken the prior’s guidance.
  2. Depth can make the latent code for the same intent vary as the robot moves. For example, when the robot gets up from the floor, depth observations can change rapidly or become noisy, making the intent harder to encode consistently.
MethodSeen Joint Position Error ↓Seen Root Velocity Error ↓Seen Root Heading Error ↓Seen FS ↓Seen RR ↑Unseen Joint Position Error ↓Unseen Root Velocity Error ↓Unseen Root Heading Error ↓Unseen FS ↓Unseen RR ↑
B with Depth0.14 ± 0.050.31 ± 0.219.4 ± 6.30.040 ± 0.02330.20.15 ± 0.060.25 ± 0.178.1 ± 5.30.028 ± 0.01650.3
P-BFM0.11 ± 0.040.25 ± 0.126.9 ± 5.40.043 ± 0.02641.20.13 ± 0.060.21 ± 0.126.7 ± 4.70.026 ± 0.01661.4

2. Simpler deployment on real robots. P-BFM lets users specify tasks through rewards or intent latents. With a terrain-agnostic backward encoder B, we can compute the intent latent z offline or provide it asynchronously. A depth-conditioned encoder would need to infer z online from the current depth observation, adding latency and complicating deployment. Differences between simulated and real depth observations would also introduce an additional sim-to-real gap in intent encoding.

We hope these clarifications and additional experiments address your concerns. We will incorporate these clarifications and additional experimental results into the revised manuscript.

Thank you for your added experiments. This additional study on terrain-agnostic intent encoding has addressed my main concern. I'm raising my rating to 4.

Reviewer_zjwnOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper propose P-BFM(perceptive Behavior Foundation model for humanoid). The key contribution is separate terrain behavioral intent and other goal intent like motion tracking, reward. The other contribution would be introducing perception into general motion tracker called BFM, a mechanism for fixing root tracking error called ERI, and good performance in experiments and sim2real. From my understanding this paper enable humanoid has multi objective and generate human-like motion with perception input for terrain traversal.

Critical strengths

Concrete problem formulation: The author propose a concrete problem formulation that current general motion tracking should consider perception as well.

Real-world experiments: The video demo is impressive and show the robustness under the real-world disturbance.

Critical weaknesses

Lack of baselines: I can understand author want to emphasize they can do a bunch of things with terrain, but the experiments are not fully convincing for me that this whole pipleline is essential.

The writing makes it looks branching and deviate from the motivation, it’s not easy for reader to understand what’s the main goal of the paper. For better terrain traversal? then author might need to show more challenging terrain like previous perceptive parkou work. For motion tracking, or for better teleoperation…? If it is then need corresponding stronger baseline and quantitative result.

Questions for authors

Lack of baselines: Why not training a general motion tracker with large randomization and perception which have to traverse terrain while maintaining human-like behavior. Also you can rl post-training a refernce motion generator condition on camera as a upstream Input for general tracker, to generate condition on either goal, velocity, heading…. like what this paper did. Personaly not sure why BFM and pipeline is necessary here.

The motivation “Humanoid robots are designed to operate in human-centric environments, with versatile embodiments suited to diverse physical tasks.” seems didn’t been demonstrated, i.e. pick and place. Does this general tracker is accurate enough in flat ground to enable pick and place task like sonic? I worried it enable more complaince in terrain traverse but scrafice the accuracy which is contradictory with its motivation.

How’s the tracking performance (Lafan) in flat ground, like body pos error .etc comparing with SOTA?

Under author’s setting, how would the robot distinguish “object” and “terrain”, what would happen if robot is holding a box while camera identify it’s a ground? The current paper and video lack this demonstration and analysis.

Motion tracking are always stand as a downstream wbc controller for upstream high-level controller like VLA, Diffusion Policy….. It seems that you can also use upstream imitation learning policy for terrain traversal right? Conversely, if the downstream wbc controller is like P_BFM which has the perception input, it could distort the high level intent, such as upstream VLA output cleaning a sofa/tatami motion, your motion tracker might want jump over it.

6 comments
AuthorsOpenReview ↗

Rebuttal: Baselines, Task Scope, and Precision (1/2)

Thank you for your thoughtful feedback and for recognizing our real-world robot results. We address your questions about our motivation, baselines, tracking accuracy, and integration with high-level controllers below. The response and additional results are also available on the anonymous rebuttal page.

W1 / Q1. Motivation for FB and Comparison with Tracking-Based Approaches

We use FB to train the policy on diverse reward-maximization objectives through interaction with the terrain. The rewards directly guide terrain adaptation, while flat-ground motion data provides a behavioral prior. Our improved performance comes from optimizing task objectives that remain valid across terrains.

Baseline 1. Flat-Ground Motion Tracking. On uneven terrain, precisely following flat-ground references can conflict with the movements needed to traverse the terrain. Relaxing the tracking objective can help the robot climb stairs, but it can also reduce motion fidelity, leading to less controlled upper-body motion and unsteady steps. We compare P-BFM with MimicLite augmented with depth observations (MimicLite + Depth).

MethodSeen Joint Pos. (rad) ↓Seen Root Vel. (m/s) ↓Seen Root Heading (°) ↓Seen FS (cm) ↓Seen RR (%) ↑Unseen Joint Pos. (rad) ↓Unseen Root Vel. (m/s) ↓Unseen Root Heading (°) ↓Unseen FS (cm) ↓Unseen RR (%) ↑
MimicLite + Depth0.16 ± 0.050.29 ± 0.2011.2 ± 15.70.159 ± 0.06728.30.17 ± 0.060.24 ± 0.139.4 ± 16.50.141 ± 0.06744.7
P-BFM0.11 ± 0.040.25 ± 0.126.9 ± 5.40.043 ± 0.02641.20.13 ± 0.060.21 ± 0.126.7 ± 4.70.026 ± 0.01661.4

[1] RoboParty Lab Team. MimicLite: Efficient and Effective General Humanoid Motion Tracking. 2026.

Baseline 2. Terrain-Compatible Reference Generation. PMT uses offline trajectory optimization, which can be cumbersome and yield suboptimal references for contact-rich behaviors. A learned reference generator would also require suitable paired data or supervision to learn these adaptations.

FB lets the policy explore directly on the terrain and learn to satisfy task objectives without generating terrain-specific references.

W2 / Q4 / Q5. Task Scope and Integration with High-Level Controllers

Our goal is to enable reliable motion tracking, goal reaching, and reward optimization on everyday terrain. With only flat-ground motion data, we aim to execute diverse behaviors accurately across terrains. We leave highly challenging settings such as parkour for future work, since ordinary walking may provide a weak prior for these tasks in our setting. Nevertheless, the approach is general and could use parkour motions as a prior to potentially generalize to obstacles of different heights.

A high-level controller specifies what P-BFM should do through a reward function or an intent latent. For example, a forward-velocity reward can guide the robot across a sofa, a motion-tracking objective can guide it to dance on the sofa, and a reward for repeated hand sweeps can guide it to clean the sofa. The high-level controller could also handle terrain adaptation, but learning to generate terrain-compatible references would require terrain-specific training data, making it harder to train. The planner would also need to update its outputs fast enough to keep up with changes in terrain. In P-BFM, the high-level controller specifies terrain-agnostic task goals, and the low-level policy adapts its actions to the observed terrain to achieve them. (Q5)

When a box is in front of the robot, a reward for crouching, reaching, and bringing its arms together can guide the robot to grasp the box. A forward-motion reward can guide it to step over the box but the robot does not step over the object by default.. Occlusion from a carried object remains an open challenge for perception-based controllers, and we leave this case for future evaluation. (Q4)

AuthorsOpenReview ↗

Rebuttal: Baselines, Task Scope, and Precision (2/2)

Q2 / Q3. Tracking Precision

We additionally compared P-BFM with UFO trained and evaluated on flat ground (Appendix A.14, Table 14, reproduced below). P-BFM achieves comparable joint-position accuracy and lower global root error with ERI. The UFO paper also reports tracking accuracy comparable to state-of-the-art trackers, with substantially greater robustness.

MethodSeen Joint Pos. (rad) ↓Seen Relative Body Pos. (cm) ↓Seen Root Vel. (m/s) ↓Seen Root Heading (°) ↓Seen FS (cm/step) ↓Unseen Joint Pos. (rad) ↓Unseen Relative Body Pos. (cm) ↓Unseen Root Vel. (m/s) ↓Unseen Root Heading (°) ↓Unseen FS (cm/step) ↓
UFO0.122 ± 0.02511.56 ± 5.620.525 ± 0.38240.88 ± 27.100.0646 ± 0.04130.133 ± 0.0487.89 ± 4.530.279 ± 0.20223.84 ± 21.300.0390 ± 0.0263
P-BFM0.104 ± 0.0364.63 ± 2.360.213 ± 0.0866.15 ± 4.740.0375 ± 0.02450.128 ± 0.0574.75 ± 2.340.178 ± 0.0806.19 ± 4.060.0211 ± 0.0130

We train P-BFM without a tracking reward. Motion tracking is one of several downstream tasks. A single robust policy can follow motion references, reach goals, and optimize newly specified rewards across terrains without task-specific retraining or terrain-specific demonstrations.

We hope these clarifications and additional experiments address your concerns. We will incorporate these clarifications and additional experimental results into the revised manuscript.

Reviewer_zjwnOpenReview ↗

Response to Author Rebuttal

Thank you for the response and the experiments. The new results address part of my concern, but I still have a few questions.

P-BFM receives reference-derived heading and planar-velocity instructions through ERI, while UFO keeps its original interface? Is that a apple2apple comparison?

For the flat-ground tracking comparison, could you also report results against MimicLite or another strong dedicated motion-tracking baseline under the same evaluation protocol? Eg. Sonic(SciRobo2026), supermimic(CORL2026), how many motions would it fail in Bone-seed?

In real-world motion tracking, how is the current heading required by ERI obtained? Does deployment require LIO, external localization, or another source of global yaw estimation?

I do not find the response to W2/Q4/Q5 fully convincing. My concern is about what could actually happen in practice. The current experiments do not yet demonstrate that “the approach is general,”. At least should not sacrifice agility. And in the sofa example, stepping or jumping over the sofa seems like a plausible behavior that could indeed emerge under the proposed framework. I don't think simply leave them to future work helps.

Reviewer_wN4XOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes P-BFM which is a terrain-aware whole-body humanoid controller using unsupervised RL that only relies on a flat-ground motion dataset. P-BFM extends the BFM-Zero/UFO framework with an established depth image encoder and terrain training. The authors explicitly condition the policy with root heading and planar velocity commands in some training and deployment tasks which they argue is necessary to prevent excessive drifting, i.e. the policy ignoring root orientation and velocity rewards to avoid challenging terrain. P-BFM’s terrain-aware motion tracking is validated in simulation against its non-perceptive predecessor UFO/BFM-Zero as well as PMT, a recent method for terrain-aware motion tracking. The authors further ablate the explicit root command conditioning, showing the increase in root drift. On real hardware, the paper presents terrain-aware goal-reaching, motion tracking, zero-shot reward execution and teleoperation from flat-ground motions.

Critical strengths

This paper represents a good empirical contribution: Scaling BFMs to terrain-aware whole-body humanoid control and doing so without the need of terrain-adjusted reference motions. They show that BFMs are competitive in this domain and which challenges can arise (ERI).

Critical weaknesses

ERI conditions the intent-conditioned policy additionally on root heading and planar velocity commands and trains it with the corresponding tracking rewards. This constitutes an extra supervision signal during training that goes beyond the regularization rewards present in prior work on scaling unsupervised RL (UFO/BFM-Zero). The experimental section should be more clear in which tasks these commands are used and when they are masked out, especially for the reward inference since there is substantial overlap between the reward functions used in the experiments and during training.

Questions for authors
  1. L299-300: “and uses ERI heading and planar-velocity instructions whenever valid”: Does this mean ERI is also used for goals and/or for the demonstrated zero-shot reward inference on velocity tracking commands? If so, how does performance change when adding/removing this explicit command?
  2. Are the heading and planar velocity commands used in training also encoded in the intent latent vector? I.e. is this information in the latent vector? If so, how well does a version of P-BFM perform that does not see the explicit command in the policy, but keeps the ERI rewards (and critics)?
  3. How does P-BFM perform on reward inference for rewards that describe more complicated behavior, such as the ones used in HumEnv [1]? Especially rewards that do not have a strong overlap with planar velocity tracking?
  4. How does P-BFM behave when the ERI command demands the root to move in a different direction than the current intent does?
  5. It is currently not clear from the text whether the UFO baseline was trained on flat ground or with the same terrain used by the other methods. Could you clarify this?
2 comments
AuthorsOpenReview ↗

Rebuttal: ERI and Baseline Comparisons (1/2)

Thank you for your thoughtful feedback and for recognizing our empirical contributions. We appreciate your insightful questions regarding the role of ERI and the baseline comparisons. We address these concerns in detail below. The additional results and videos are also available on the anonymous rebuttal page.

W1 / Q1. When and How Is ERI Used?

ERI primarily serves two purposes during training: (1) guiding policy exploration toward challenging motions on diverse terrains and (2) enabling the policy to explicitly condition on root-level commands.

During inference, the use of ERI depends on the downstream task:

  • Motion tracking. We extract root-level commands from the reference motion and provide them as ERI inputs.
  • Goal inference. ERI is disabled.
  • Reward inference. ERI is optional. For the forward-velocity task, we evaluate both ERI-assisted inference and pure reward inference with both ERI inputs disabled. All posture-based reward examples use pure reward inference.

Both motion tracking and reward inference can still be performed without explicit ERI inputs. However, removing ERI reduces the policy’s ability to follow global root-level commands, as shown below. Watch the motion-tracking comparison.

ERISeen EjposE_{\mathrm{jpos}} (rad) ↓Seen EvelrootE_{\mathrm{vel}}^{\mathrm{root}} (m/s) ↓Seen EheadrootE_{\mathrm{head}}^{\mathrm{root}} (°) ↓Seen FS (cm) ↓Unseen EjposE_{\mathrm{jpos}} (rad) ↓Unseen EvelrootE_{\mathrm{vel}}^{\mathrm{root}} (m/s) ↓Unseen EheadrootE_{\mathrm{head}}^{\mathrm{root}} (°) ↓Unseen FS (cm) ↓
On0.11 ± 0.040.25 ± 0.126.89 ± 5.370.043 ± 0.0260.13 ± 0.060.21 ± 0.126.74 ± 4.650.026 ± 0.016
Off0.11 ± 0.040.53 ± 0.3836.77 ± 25.510.029 ± 0.0160.13 ± 0.050.37 ± 0.2526.70 ± 23.790.018 ± 0.010

In our 0.8 m/s forward-motion reward-inference tests, disabling ERI increased the mean velocity error from 0.132 to 0.272 m/s.Watch the reward-inference comparison.

Q2. Overlap Between the Inputs to B and ERI Commands

During training, the backward representation BB encodes the state into a goal latent. Since the state contains root-level information such as base velocity, it partially overlaps with the information provided by ERI. This intentional redundancy allows the policy to follow the goals specified by BB even without explicit ERI commands, as in certain reward inference tasks. We also trained a model without explicit root commands as actor inputs (Actor w/o ERI) and a model that removes the overlapping information from the inputs to BB (B w/o ERI). The results are shown below.

MethodSeen Eˉjpos\bar{E}_{\mathrm{jpos}} ↓Seen EvelrootE_{\mathrm{vel}}^{\mathrm{root}} ↓Seen EheadrootE_{\mathrm{head}}^{\mathrm{root}} ↓Seen FS ↓Seen RR ↑Unseen Eˉjpos\bar{E}_{\mathrm{jpos}} ↓Unseen EvelrootE_{\mathrm{vel}}^{\mathrm{root}} ↓Unseen EheadrootE_{\mathrm{head}}^{\mathrm{root}} ↓Unseen FS ↓Unseen RR ↑
Actor w/o ERI0.12 ± 0.040.39 ± 0.2423.4 ± 16.70.041 ± 0.02311.10.12 ± 0.060.35 ± 0.2124.4 ± 21.10.021 ± 0.01317.1
B w/o ERI0.12 ± 0.040.33 ± 0.237.3 ± 6.40.041 ± 0.02431.80.13 ± 0.060.27 ± 0.195.8 ± 5.30.025 ± 0.01654.0
P-BFM0.11 ± 0.040.25 ± 0.126.9 ± 5.40.043 ± 0.02641.20.13 ± 0.060.21 ± 0.126.7 ± 4.70.026 ± 0.01661.4

Watch the ablation comparison.

W2 / Q3. Reward Inference Without ERI Commands

We have added reward-inference examples of single-leg balance, asymmetric arm postures, and crouching with both hands reaching forward, as shown in the video. All examples use the same frozen policy with ERI disabled, without additional training. These examples demonstrate behavior generation beyond velocity tracking, although precisely satisfying complex combinations of objectives remains challenging. Watch the reward-inference examples.

AuthorsOpenReview ↗

Rebuttal: ERI and Baseline Comparisons (2/2)

Q4. Conflicts Between Behavioral Intent and ERI Commands

The FB objective, defined through QFBQ_{\mathrm{FB}}, aligns the policy’s behavior with the goal latent produced by BB, while ERI encourages explicit root-command following. When the goal state conflicts with the ERI command, the two Q-value terms jointly influence policy optimization according to their respective weights. A larger ERI weight prioritizes command following, whereas a smaller weight places greater emphasis on the FB objective.

We evaluated our final trained model by keeping the policy and forward-walking intent fixed while varying only the ERI commands. As shown in the video, rotating both the heading and velocity targets in ERI by 90° redirects the robot’s motion accordingly, demonstrating consistent directional retargeting. Keeping the heading target unchanged while reversing only the velocity target causes the robot to slow down substantially and step in place, without reliably moving backward.

These results indicate that ERI can effectively guide behavior but does not take absolute priority over the behavioral intent. While the current method does not guarantee reliable execution under arbitrary conflicting intent–command combinations, the relative priority of ERI can be adjusted by tuning the objective weights when stronger command following is desired.

Watch how the policy responds to each command.

Q5. UFO Baseline

The UFO baseline reported in our paper was trained on flat ground. We additionally evaluated a new UFO baseline trained on diverse terrains, with the results presented below.

MethodSeen EjposE_{\mathrm{jpos}} (rad) ↓Seen EvelrootE_{\mathrm{vel}}^{\mathrm{root}} (m/s) ↓Seen EheadrootE_{\mathrm{head}}^{\mathrm{root}} (°) ↓Seen FS (cm) ↓Seen RR (%) ↑Unseen EjposE_{\mathrm{jpos}} (rad) ↓Unseen EvelrootE_{\mathrm{vel}}^{\mathrm{root}} (m/s) ↓Unseen EheadrootE_{\mathrm{head}}^{\mathrm{root}} (°) ↓Unseen FS (cm) ↓Unseen RR (%) ↑
UFO0.15 ± 0.060.65 ± 0.4649.6 ± 29.30.092 ± 0.0503.40.16 ± 0.060.37 ± 0.2728.9 ± 23.90.066 ± 0.04716.6
UFO-Terrain0.12 ± 0.040.45 ± 0.3431.4 ± 22.40.090 ± 0.0409.70.14 ± 0.060.29 ± 0.2123.0 ± 20.40.061 ± 0.03423.9
P-BFM0.11 ± 0.040.25 ± 0.126.9 ± 5.40.043 ± 0.02641.20.13 ± 0.060.21 ± 0.126.7 ± 4.70.026 ± 0.01661.4

We hope these clarifications and additional experiments address your concerns. We will incorporate these clarifications and additional experimental results into the revised manuscript.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

We thank the reviewers for their constructive feedback. We are glad they recognized the value of terrain-aware BFMs without terrain-matched demonstrations (wN4X, JCBi), the clear problem formulation and robust real-world results (zjwn), and the separation of intent and execution and the ERI design (JCBi).

All three current ratings are positive (3, 3, and 4). Our additional experiments and videos are available on the anonymous rebuttal website.

  1. Reviewer wN4X — Additional ERI supervision and reward inference. We clarified ERI usage across tasks and added ablations of command inputs during training and ERI on/off at inference. These experiments clarify ERI’s role in guiding exploration and improving root-control precision. We also added more diverse reward-inference examples with ERI commands disabled, demonstrating that the same frozen policy can realize varied posture and balance objectives. Results

  2. Reviewer zjwn — Framework justification and stronger baselines. We explained the motivation for learning terrain adaptation through unsupervised exploration when terrain-matched references are unavailable, and added supervised tracking baselines. The comparisons show improved terrain tracking and recovery, alongside flat-ground global tracking accuracy comparable to a dedicated tracker. These results support P-BFM’s value as a general controller across terrains. The reviewer subsequently confirmed their positive rating. Results, additional comparisons

  3. Reviewer JCBi — Value of terrain-agnostic intent encoding. We added a direct ablation that supplies depth observations to the backward encoder. Excluding depth improves tracking accuracy and recovery, directly supporting the separation of intent encoding from terrain-adaptive execution. The reviewer confirmed that this experiment addressed their main concern and raised their rating to 4. Results

We have incorporated these additions into the revised manuscript. Thanks again for the constructive feedback.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.