acceptodds
Withdrawn from ICLR 2027

Beyond Joint Angles: Learning Spatial Robot Arm Structure for Collision-Free Policy

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Collision avoidance is a key challenge in bimanual manipulation. We argue that the essence of this problem is that proprioceptive representations such as joint angles lack information about the manipulator’s spatial structure. Based on this insight, we propose RoboCloud, a structural-prior descriptor that models both the geometry of entire manipulator and the topology of link connection, providing both spatial structure and kinematics for predicting potential surface contact either between the manipulators or between a manipulator and an observed object. Building on this representation, we introduce SPACE Policy, an imitation learning framework that uses Gated Bimanual Fusion to perceive spatial relationships between manipulators for self-collision avoidance and Gated Robot-Prior Fusion to perceive object–robot spatial relationships for environmental collision avoidance. SPACE Policy achieves an average success rate of 69.1% across 31 tasks on the RoboTwin 2.0 benchmark with a 11.9% improvement, and an average success rate of 83% across 7 tasks on the RLBench2 benchmark with a 2.2% improvement, and it demonstrates strong collision avoidance capability on real-world tasks. Our code will be released publicly to facilitate reproducibility and further research.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Control Policy, Robot Learning, Embodied Intelligence

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_E3ZGOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This work models a bimanual robot using point clouds. It then establishes contact relationships between the two robotic arms, as well as between the arms and the contacted objects, to achieve collision avoidance. The method is simple and effective, and the paper is very clearly written, as discussed in the first point of the Strengths section. However, the weaknesses are also obvious. The experimental design is unreasonable, and there is no cost analysis, which makes this work incomplete. In addition, the novelty is limited.

Critical strengths

The paper is quite readable. The writing is clear, and the logic flows smoothly. It is easy for readers to understand, and there are no apparent signs of AI-assisted writing. I can see that the authors have put large efforts into writing the manuscript. In particular, the narrative logic in Sections 1 (Introduction) and 2 (Related Work) is very rigorous. Figures 1 to 3 appear to be hand-drawn. Although they are not as polished or visually appealing as AI-assisted figures, they convey the ideas very clearly. There is no unnecessary or irrelevant content, allowing readers to easily understand the authors' method. The method section is not packaged with complicated formulas. It is written very concisely, and the explanations of each formula are sufficient. In addition, the design of the method's architecture is also simple. This may indicate that the authors' method is simple and effective.

Critical weaknesses

1.The experimental design is unreasonable, and the real-robot experiments are too simple.

2.The evaluation metrics are not reasonably designed. Since the work focuses on collisions, it is difficult to analyze the problem when all the metrics are task success rates. The paper also lacks related analysis.

3.The approach used for collision detection is actually already quite mature. I do not think the novelty is sufficient. In addition, if there is only a single joint angle, is a topological point cloud really necessary?

4.The experimental section lacks a cost analysis of the architecture.

Questions for authors

1.This work is somewhat similar to Motion Policy Networks (Adam Fishman et al., CoRL 2023). What is the difference between this work and the idea of using scene point clouds to learn collision-free motions in that paper?

The following are some minor suggestions, not questions:

1.In line 160, "bimanual-spatial feature" is used, but Figure 2 uses "bi-spatial feature". I suggest using consistent terminology to make it clearer for readers.

2.There is no need to include the results for all categories in Table 1. I also suggest including the publication years of the compared methods in the table. This would allow readers to directly see the relationship between how recent the methods are and their performance.

3.The title of this work contains "Beyond Joint Angles". Although this is correct, I do not think it summarizes the motivation of this work very well. It might be more appropriate to remove this phrase or revise the title. Of course, I am willing to hear the authors' rebuttal.

Reviewer_KG7cOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper argues that bimanual imitation policies collide because joint-angle proprioception does not encode the robot's spatial extent. It proposes RoboCloud, a deterministic robot descriptor computed from the URDF and the current joint angles. RoboCloud has two parts: a geometric cloud, and a set of topological clouds. Each arm's RoboCloud is encoded with PointNets and a Transformer into a global feature. The two arms are combined by a sigmoid-gated residual fusion, and the result is injected in the same way into the scene point-cloud and state features of a DP3-style diffusion policy. The paper reports 69.1% on 31 RoboTwin 2.0 tasks (three seeds), compared with 57.2% for GAP and 50.8% for DP3; 83.0% on 7 RLBench2 tasks, compared with 80.8% for PPI; and 63.2% on five real-world bimanual tasks with 100 trials each, compared with 48.0% for GAP. The supplementary material contains code and five real-robot videos.

Critical strengths
  1. A simple, cheap robot representation with large and well-sampled gains. RoboCloud needs only the URDF and forward kinematics, adds no sensors, and can be plugged into any point-cloud policy. On RoboTwin, the evaluation covers 31 tasks with 100 rollouts and three training seeds. SPACE improves on DP3 by 18.3 points and on GAP by 11.9 points, far beyond the per-task seed deviations. The implementation is practical and clearly described: sampling templates are fixed offline, and forward kinematics is applied to raw joint angles rather than normalized ones. The released code covers the RoboTwin pipeline, including the configurations for all reported ablations and the standard final-checkpoint evaluation. Together with the per-task ablation tables, this makes the main results checkable.
Critical weaknesses
  1. The collision-avoidance claim, central to the title, is not supported.

    • No collision rate is reported. App. D says that real-world collisions were only "qualitatively inspected", and Fig. 4 and the videos show only successful rollouts. The training objective has no collision-related term, and the per-task gains are broad rather than collision-specific.

    • The motivation is also overstated. The introduction argues that kinematic information alone is insufficient to determine whether a collision occurs. Together with the URDF, however, the joint angles determine the robot's full geometry. This is how collision checkers and motion planners work, and how planners produce collision-free expert demonstrations such as those in RoboTwin. The open question is therefore whether an end-to-end policy can exploit information it already has, not whether that information is missing.

    • Sec. 2.2 likewise states that existing methods "cannot handle" contact at the level of link surfaces, but it reviews only recent bimanual imitation-learning work. Representing the robot by forward-kinematics point samples is established for learning collision-free motion, in MπNets and Neural MP, and the Kinematics-Aware Diffusion Policy represents the arm as points on its body, aligned with the scene cloud. Neither these works nor the motion-planning, collision-free motion generation and safety-filtering literature is discussed.

    • As a result, the paper does not explain why learning collision awareness is preferable to collision checking or safety filtering, which use the same URDF meshes.

  2. The gain is not yet attributed to the robot's spatial structure. RoboCloud is a deterministic function of the joint angles, which the policy already receives, so it adds an inductive bias rather than information. Sec. 4.2's statement that joint angles "cannot represent the gripper's spatial position" is not accurate, since forward kinematics determines it uniquely. The ablations do not isolate spatial occupancy. Each topological cloud depends on a single joint angle and carries no world-frame occupancy, yet topology-only RoboCloud reaches 67.6%, close to 69.7% for the full model. Every variant of the robot branch scores between 63.3% and 69.7%. The "neither" row (50.8%) equals DP3's three-seed result, so it is unclear whether it is a run of the SPACE architecture without the robot branch. Moreover, robot and scene features are pooled into global vectors before a channel-wise gate, which leaves no point-level interaction from which clearances could be computed. The gain may therefore come from a richer encoding of proprioception or from added capacity.

[1] Fishman et al., "Motion Policy Networks" (MπNets), CoRL 2022;

[2] Dalal et al., "Neural MP: A Generalist Neural Motion Planner", IROS 2025.

[3] Sundaralingam et al., "cuRobo: Parallelized Collision-Free Robot Motion Generation", ICRA 2023;

[4] Jung et al., "RAIL: Reachability-Aided Imitation Learning for Safe Policy Execution", ICRA 2025;

[5] Römer et al., "From Demonstrations to Safe Deployment: Path-Consistent Safety Filtering for Diffusion Policies", ICRA 2026.

Questions for authors

A. Other Requests that could change my rating:

  1. Please report two controls on RoboTwin. The first is the SPACE architecture without the robot branch, with the scene encoder, state encoder and schedule kept unchanged. Please also confirm whether the scene encoder matches DP3's, since its output dimension appears to differ (128 vs. 64). The second is a capacity-matched control that uses the same robot branch but is fed with joint angles, for example per-joint tokens, FK link poses, or an MLP over the ordered FK keypoints.

  2. Please report simulator-measured collision rates, separated into arm–arm and arm–object contacts, for SPACE, DP3 and GAP on RoboTwin. Please also break down the success gain by task type, single-arm vs. dual-arm. If the gain is concentrated where collisions occur and collisions drop clearly, The Weakness could be resolved. Otherwise, the paper should be reframed as spatial proprioception for manipulation, with the title and claims revised accordingly.

B. Just Questions:

  1. On the real robot, DP3 reaches 12.8% while SPACE reaches 63.2%, a much larger gap than in simulation. Which backbone was used for SPACE, and were DP3 and SPACE trained with the same data and schedule? How many demonstrations were collected per task, and how were initial conditions randomized? How were the keyframe-based baselines (PPI, KStar Diffuser) adapted to 30 Hz joint-space control?

  2. Does the scene point cloud already contain the robot's points? If so, what does RoboCloud add beyond completing occluded parts of the arm?

  3. How does RoboCloud differ from the robot-point representations used in MπNets (CoRL 2022), Neural MP (IROS 2025) and the Kinematics-Aware Diffusion Policy (arXiv 2025)? When does a learned representation help beyond analytic collision checking, or beyond a safety filter on top of the policy such as RAIL (ICRA 2025)?

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

This paper proposes SPACE Policy, an imitation learning framework that incorporates explicit robot spatial structure to improve collision avoidance in bimanual manipulation. The key component is RoboCloud, a structured point cloud representation constructed from the robot's URDF and joint states, encoding both link surface geometry and kinematic topology. The framework integrates RoboCloud features with visual observations and proprioceptive states through gated fusion modules, and employs a diffusion policy for action generation. The method is evaluated on 38 simulation tasks across RoboTwin 2.0 and RLBench2, as well as five real-world bimanual manipulation tasks, demonstrating improvements in task success rates over existing baselines.

Critical strengths
  1. Practical and intuitive robot spatial representation. The paper addresses an important limitation of conventional proprioceptive representations in robotic manipulation. By explicitly incorporating link geometry and kinematic structure through RoboCloud, the policy can access information about the robot's spatial occupancy that is not directly represented by joint angles or end-effector poses. The representation is constructed from readily available URDF models and forward kinematics, making it practical to implement without additional annotations or sensors.

  2. Extensive simulation and real-world evaluations. The authors evaluate SPACE Policy on 38 simulation tasks across two benchmarks and five real-world bimanual manipulation tasks. The method achieves notable improvements over strong baselines, particularly on RoboTwin 2.0 and real-world tasks. The paper also includes ablations on the geometric and topological components, fusion strategies, point sampling configurations, and point cloud encoders.

Critical weaknesses
  1. Limited methodological novelty and insufficient justification for the proposed representation. The central idea of incorporating explicit robot geometry into policy observations is reasonable, but the technical implementation largely combines existing components, including forward kinematics, point cloud encoders, gated fusion, and diffusion policies. More importantly, the experiments do not sufficiently establish whether the proposed RoboCloud representation offers meaningful advantages over simpler geometry-aware alternatives, such as FK-based link poses, sampled robot keypoints, or capsule-based representations. Additionally, Table 4 shows that using only the geometric point cloud already achieves a 68.5% success rate, compared with 69.7% for the full RoboCloud representation. The relatively small improvement raises questions about the necessity of the additional topological representation and the complexity it introduces.

  2. Lack of direct quantitative evaluation of collision avoidance. Collision avoidance is the primary motivation and claimed contribution of this paper. However, both simulation and real-world evaluations mainly rely on task success rates. While the proposed method achieves higher success rates, these improvements do not necessarily demonstrate better collision avoidance, as they may also result from improved state representations or action prediction. The paper does not provide quantitative comparisons of self-collision rates, environmental collision rates, or minimum collision clearance. In particular, although the real-world experiments include collision-prone tasks, no quantitative collision statistics are reported. This makes it difficult to determine to what extent the observed performance improvements are actually attributable to collision avoidance rather than general improvements in manipulation performance.

Questions for authors
  1. Could the authors provide comparisons with simpler geometry-aware representations, such as FK-based link poses, sampled robot keypoints, or capsule-based geometry? Such comparisons would help clarify whether the performance gains arise from the specific RoboCloud design or simply from providing explicit robot geometry to the policy.

  2. Could the authors report quantitative collision avoidance metrics, such as self-collision rate and environmental collision rate, for both simulation and real-world experiments? Since collision avoidance is the central claim, these results would be important for supporting the paper's conclusions.

  3. In Table 4, the geometric-only variant achieves 68.5% success, while the full method achieves 69.7%. Could the authors further explain or demonstrate scenarios where the topological representation provides substantial benefits beyond the geometric point cloud?

  4. Could the authors clarify the demonstration collection and training protocols for the five real-world tasks, including the number of demonstrations per task and whether all baselines use the same training data?

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

Related papers

Discussion (0)

Sign in to comment.