acceptodds
Under review as a conference paper at ICLR 2027

BARS-RL: Bimanual Articulated-object manipulation under Resistance with Stabilization

Abstract

Robots in human environments frequently interact with articulated objects that rest unanchored on a surface (e.g., microwaves, toaster ovens), and opening these objects requires overcoming an initial opening resistance. A pulling force below this resistance leaves the moving part closed, whereas one above it can drag the object body, so one arm must pull while the other holds the object body in place. Existing methods that assume anchored objects or negligible opening resistance do not capture this coupling between opening and stabilization. To address these limitations, we propose BARS-RL, a reinforcement learning framework in which a single policy allocates force between a task arm and a support arm without observing the resistance or friction. A stabilization reward penalizes contact forces near the base-slip boundary, providing a training signal before the object body moves. Training follows a curriculum organized by the resistance-to-friction ratio, which moves from an anchored object body to an unanchored one and finally mixes episodes that base friction alone can hold with those that require support. On three simulated objects with revolute and prismatic joints, BARS-RL achieves an average success rate of 83.5%, outperforming baselines while approaching the performance of an oracle that observes the resistance and friction.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Reinforcement Learning, Articulated Object Manipualtion, Bimanual Manipulation, Curriculum Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_mRm6OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

In human environments, robots often need to operate articulated objects such as microwaves and toaster ovens that rest unanchored on a surface. Opening them requires overcoming an initial opening resistance: too little pulling force leaves the door or lid closed, while too much can drag the entire object, so one arm typically pulls while the other stabilizes the object body. Existing methods either assume the object is anchored or ignore opening resistance, and thus fail to capture this coupling between opening and stabilization. To address this, the authors propose BARS-RL, a reinforcement learning framework in which a single policy allocates force between a task arm and a support arm without observing resistance or friction. Its key design is a stabilization reward that penalizes contact forces near the base-slip boundary, providing a training signal before the object body moves; training follows a curriculum organized by the resistance-to-friction ratio, progressing from an anchored object body, to an unanchored one, and finally to a mix of episodes that base friction alone can hold and those that require support.

Critical strengths
  1. The core physical quantity p can explain when a support arm is needed, and the ablation studies also confirm this phenomenon.
  2. The problem is of practical significance: for articulated objects that are not anchored in the environment, such as microwaves, refrigerators, and air fryers.
  3. The paper is well structured, clear, and easy to understand.
Critical weaknesses
  1. Currently, the results are entirely simulation-based, lacking real-robot experiments.
  2. The baselines are relatively simple, lacking comparison with stronger recent single-arm/dual-arm methods.
  3. The reward function contains many handcrafted weights, thresholds, and safety coefficients, but no sufficient sensitivity analysis has been conducted.
Questions for authors
  1. The simulation scenarios addressed by this method are rather idealized, making it difficult to apply them to real robots.
  2. The baselines selected are relatively few and do not adequately demonstrate the advantages of the method.
  3. Do the maximum force exerted by the support arm and the impact experienced by the object satisfy practical safety constraints?
  4. Is the current method sensitive to the weights of the multiple reward functions?
  5. How can it be explained that the version without curriculum achieves 0% on all objects?
  6. Are the support arms intentionally set in advance to keep each object in a fixed end-effector pose? It seems like the fixed end-effector poses might be different for manipulating different objects?
8 comments
AuthorsOpenReview ↗

Response to Reviewer mRm6 (1/3)

We thank the reviewer for recognizing that the ratio ρ explains when a support arm is needed and that the problem has practical significance. The reviewer's concerns were the lack of real-robot experiments, the baselines, and reward sensitivity. We have addressed each with new experiments; all changes in the revised PDF are marked in red.

Summary of changes

  • Real robots (W1, Q1): 19/20 real-world successes, versus 0/20 for single-arm RL; real closure forces match the breakaway model (Sec. 5.2, Fig. 6).
  • Baselines (W2, Q2): ACT, Diffusion Policy, and VoxAct-B, each ± F/T, added; the best reaches 45.1% vs. 83.5% (Table 1).
  • Reward sensitivity (W3, Q4): sweep of the base-slip cost parameters, 57.2–69.0% (Table 6); one reward setting reused across five objects and two robots.
  • Safety, curriculum, initialization (Q3, Q5, Q6): force logs from real trials, a training-failure analysis, and automatic pose selection (App. E, C.1, D.1).

What we see as the main insight is that a single quantity, ρ, predicts when support becomes necessary across objects with different mass, friction, and joint type (Figure 7), and organizing training by ρ outperforms organizing it by resistance alone (Table 3).

W1 / Q1:

Currently, the results are entirely simulation-based, lacking real-robot experiments. / The simulation scenarios addressed by this method are rather idealized, making it difficult to apply them to real robots.

Response:

  • Real-world deployment (Section 5.2, Table 2, Figure 5): Simulation-trained BARS-RL succeeded in 9 of 10 microwave trials and all 10 air-fryer trials, whereas single-arm RL dragged the body and failed in all 20 trials.
  • Closure-model validation (Section 5.3, Figure 6): Joint-torque-based measurements on three real appliances showed force buildup and a sharp drop at release, matching the breakaway model, including the magnetically latched mini fridge.

W2 / Q2:

The baselines are relatively simple, lacking comparison with stronger recent single-arm/dual-arm methods. / The baselines selected are relatively few and do not adequately demonstrate the advantages of the method.

Response:

To our knowledge, no prior RL method addresses this coupled setting, so we added ACT, Diffusion Policy, and the acting-and-stabilizing bimanual method VoxAct-B, each with and without wrist F/T (Table 1; details in Appendix D.7). Average success (%):

Methodρ < 1.25ρ ≥ 1.25
ACT w/o F/T31.918.3
ACT w/ F/T35.316.8
Diffusion Policy w/o F/T41.821.9
Diffusion Policy w/ F/T64.237.7
VoxAct-B w/o F/T38.222.4
VoxAct-B w/ F/T58.345.1
Fixed-force bimanual controller39.923.8
Single-arm RL97.029.5
BARS-RL90.183.5
Oracle-physics actor (ζ observed)93.090.0

For ρ ≥ 1.25, BARS-RL exceeds the best baseline by 38.4 points. Single-arm RL leads for ρ < 1.25, where one arm suffices, but drops to 29.5% when the body is dragged before release.

(Continued in the next comment.)

AuthorsOpenReview ↗

Response to Reviewer mRm6 (2/3)

W3 / Q4:

The reward function contains many handcrafted weights, thresholds, and safety coefficients, but no sufficient sensitivity analysis has been conducted. / Is the current method sensitive to the weights of the multiple reward functions?

Response:

We agree this needed to be examined. We varied the weight and safety factor of the base-slip cost, the term we introduce, and tested whether one reward setting transfers across objects and robots.

  • Parameter sensitivity (Appendix C.3, Table 6): We used the mini fridge, the object with the lowest success rate and the largest drop without cbasec_\text{base}, and retrained with one parameter changed at a time (three seeds, ρ ∈ [1.25, 2.0]). Because these are separate runs, the default differs from Table 1 (62.2% vs. 68.4%).

    SettingSuccess rate
    Default (λbase\lambda_\text{base} = 0.04, κ\kappa = 0.7)62.2 ± 6.1
    λbase\lambda_\text{base} × 0.559.4 ± 14.6
    λbase\lambda_\text{base} × 257.2 ± 11.2
    κ\kappa = 0.569.0 ± 4.5
    κ\kappa = 0.862.6 ± 3.6

    Three of the four variants are within 5.0 points of the default, which is within its seed standard deviation (6.1). κ=0.5\kappa = 0.5 is 6.8 points higher, so the default is not tuned to the best value; all reported results use the default. Changing λbase\lambda_\text{base} mainly increases seed-to-seed variability.

  • Reuse across objects and robots: The same weights and thresholds (Tables 10 and 11) were used for all five simulated objects without per-object tuning, and the real-world policies used the same reward with only the robot model replaced (Section 5.2).

Q3:

Do the maximum force exerted by the support arm and the impact experienced by the object satisfy practical safety constraints?

Response:

The setup limits loads through a force-command limit, torque-aware initialization, and distributed contact. The commanded pressing force is limited to 40 N in training and deployment (Appendix D.1), and the initialization selects a support pose whose joint torques stay below 90% of their limits under this load. Reward penalties discourage excess grip and off-axis force, and the four flat fingers spread the support load.

Force logs were available for 13 real-world trials (10 microwave and 3 air fryer; Appendix E). The commanded pressing force stayed within 40 N. The force estimated from joint torques averaged 13–31 N per trial and peaked at 48.4 N, exceeding 40 N for at most 0.33 s, and the 80 N safety stop was never triggered. The commanded force dropped to about 5 N after latch release, and no appliance damage was observed.

These estimates are logged at 30 Hz and do not resolve millisecond-scale impact peaks, so we do not claim a direct measurement of impacts.

(Continued in the next comment.)

AuthorsOpenReview ↗

Response to Reviewer mRm6 (3/3)

Q5:

How can it be explained that the version without curriculum achieves 0% on all objects?

Response:

We added a termination breakdown of training without the curriculum on the mini fridge (Appendix C.1, Figure 8). Success stays at zero for all plotted seeds, and each seed fails in a distinct way: excessive body displacement (seed 6), time-outs (seed 5), or an increasing fraction of simulation divergence (seed 4). In the last case, since every step incurs a time penalty and the success bonus is never obtained, ending an episode early yields a higher return, and the policy learns to terminate early.

Training directly on the final distribution requires discovering pulling and pressing at the same time; until both occur, pulling only drags the body and the closure never releases, so no success signal arrives. The curriculum separates the skills: anchored opening, then unanchored opening that base friction can support, then support-dependent episodes (Section 4.3), with success-gated stage transitions (Table 13).

Q6:

Are the support arms intentionally set in advance to keep each object in a fixed end-effector pose? It seems like the fixed end-effector poses might be different for manipulating different objects?

Response:

The initial support pose is now selected by a geometry-based algorithm with common rules and parameters across objects and robots, without manual specification (Appendix D.1). It combines analytical reachability and force-feasibility analysis with geometric checks, so the pose adapts to each object's geometry.

This pose defines only the starting contact. The support hand starts without pressing, and the policy then commands both a relative pose and a wrench target (Appendix D.2), so the stabilization location and force are controlled by the learned policy through feedback.

We hope these results address your concerns. If they resolve the questions about real-robot validation, baselines, and sensitivity, we would be grateful if you would reconsider your assessment. If any concern remains, please let us know and we will address it during the discussion period.

Reviewer_NaGfOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper designs an RL-training curriculum, called BARS-RL, for a practical situation, where opening a lightweight appliance and move the whole appliance before its door open. BARS-RL is specifically designed for training a robot arm to pull while the other arm presses down to keep the object steady. The main contribution is that BARS-RL partition the training in three phases, gradually attacking the problems from purely anchored to full resistence-to-friction ratio ρ\rho, and use ρ\rho to determine if there is a need for two hands to get involved. The paper benchmark the result with fixed-force bimanual manipulation, single arm manipulation, fixed curriculum BARS-RL and curriculum-based BARS-RL, and shows promising success rate.

Critical strengths
  1. The central problem that this paper tries to address is meaningful, that is, the coupling between the opening behavior and the stabilization of the appliance. Explicitly studying how two arms share forces in this setting extends the usual articulated-object manipulation and identifies an actual challenge for real-world application.
  2. The proposed method explicitly breaks a complex problem into three stages so that RL can gradually address a more and more complexed problem with good "prior". Meanwhile, training is grounded in task's physics and results show meaningful improvements compared to the baselines.
Critical weaknesses
  1. The contribution of this paper is mainly a task-specific integration of existing/established techniques. The approach combines asymmetric actor-critic learning, hybrid force/position control, reward shaping, and curriculum learning. It's most distinctive contributions are the resistance-to-friction ratio ρ\rho and the associated curriculum, derived just from a really simplified friction model. Though the method is physically motivated, the experiments do not show the benefit of how this specific setting is better than other existing bimanual techniques and designs. Hence, this makes the paper's advantages less clear.
  2. The validation is too narrow to establish practical generality. The experiments train policy for three objects respectively with different prescribed threshold-based closure model and privileged physical information (friction etc). Since there is no real-world example, it is hard to tell if the method can be generalized well, and how accurate should the prescribed closure model actually be.
Questions for authors
  1. How is each episode initialized? For the holding hand, do you assume that the robot arm is already at a grasping position? Do you assume that the support arm is always applying a downward force at a fixed location with the same hand gesture?
  2. What explains the failure of BARS-RL without curriculum-learning? How does the learning curves look like? Are baselines not converging, plateaued, or broken? Clarifying the training trajectories of both variants, especially the learning curves and immediate success rates, can help substantiate the contribution.
  3. What evidence isolates the contribution of slip-margin reward?
5 comments
AuthorsOpenReview ↗

Response to Reviewer NaGf (1/3)

We thank the reviewer for recognizing that the coupling between opening and stabilization is a meaningful real-world challenge and that our staged, physics-grounded training improves on the baselines. The reviewer's concerns were novelty and the breadth of validation. We address both below; all changes in the revised PDF are marked in red.

Summary of changes

  • Novelty (W1): the ρ-based formulation predicts when support is needed (Fig. 7), and its two derived components add 15 and 24 points over generic alternatives (Tables 3, 4).
  • Comparison (W1): ACT, Diffusion Policy, and VoxAct-B, each ± F/T, added; the best reaches 45.1% vs. 83.5% (Table 1).
  • Validation (W2): 19/20 real-world successes, real closure measurements, gradual-release tests, and two new objects (Secs. 5.2–5.3).

W1:

The contribution of this paper is mainly a task-specific integration of existing/established techniques. [...] Though the method is physically motivated, the experiments do not show the benefit of how this specific setting is better than other existing bimanual techniques and designs.

Response:

We agree that asymmetric actor–critic, hybrid force/position control, and curriculum learning are established tools, and we do not claim them as contributions. Our contribution is the formulation that determines how they are used: from the base-slip margin (Eq. 1) we derive the resistance-to-friction ratio ρ, and this yields components that generic counterparts do not provide.

  • ρ predicts when support is needed (Figure 7). Across objects with different mass, friction, and joint type, single-arm success collapses between ρ = 1.25 and 1.5 (11% on the microwave and 0% on the others at ρ = 1.5), and all 20 real-world single-arm trials fail by dragging the body. Despite its simplifications, the model identifies the regime transition.
  • Pre-slip stabilization (Table 4). cbasec_\text{base} penalizes force states approaching slip before the body moves. With the generic displacement cost alone, success drops from 83.5% to 59.9%.
  • Ordering by ρ, not by difficulty (Table 3). With everything else identical, scheduling by FresF_\text{res} instead of ρ lowers success from 83.5% to 68.6%, because the same FresF_\text{res} needs different support under different mass and friction.

These comparisons share BARS-RL's observations and training, so they measure the effect of the formulation itself.

Comparison with existing methods (Table 1; Appendix D.7). To our knowledge, no prior RL method addresses this coupled setting, so we evaluated learning-based alternatives, including the acting-and-stabilizing bimanual method VoxAct-B, each with and without wrist F/T. Implementation details, including each method's inputs, are given in Appendix D.7. Average success (%):

Methodρ < 1.25ρ ≥ 1.25
ACT w/o F/T31.918.3
ACT w/ F/T35.316.8
Diffusion Policy w/o F/T41.821.9
Diffusion Policy w/ F/T64.237.7
VoxAct-B w/o F/T38.222.4
VoxAct-B w/ F/T58.345.1
Fixed-force bimanual controller39.923.8
Single-arm RL97.029.5
BARS-RL90.183.5
Oracle-physics actor (ζ observed)93.090.0

For ρ ≥ 1.25, BARS-RL exceeds the best baseline by 38.4 points and reaches about 93% of the oracle.

(Continued in the next comment.)

AuthorsOpenReview ↗

Response to Reviewer NaGf (2/3)

W2:

The validation is too narrow to establish practical generality. The experiments train policy for three objects respectively with different prescribed threshold-based closure model and privileged physical information (friction etc). Since there is no real-world example, it is hard to tell if the method can be generalized well, and how accurate should the prescribed closure model actually be.

Response:

We agree that broader validation was needed. We added real-world experiments, tested the closure model against measurements and against gradual release, and added new object geometries. We also clarify which information the actor receives.

  • Actor inputs: The actor does not observe ζ; these parameters are given only to the asymmetric critic during training. The actor acts from proprioception, wrist F/T, tactile signals, and object state (Table 8). Table 1 separately reports an oracle-physics actor that observes ζ (90.0% vs. 83.5% for ρ ≥ 1.25).

  • Real-world applicability (Section 5.2, Table 2, Figure 5): Simulation-trained policies succeeded in 9 of 10 microwave trials and all 10 air-fryer trials, whereas single-arm RL failed in all 20 trials by dragging the body.

  • Closure measurements (Section 5.3, Figure 6): Pull forces measured on three real appliances and simulated with the measured breakaway forces both showed force buildup followed by a sharp drop at release, including for the magnetically latched mini fridge.

  • Gradual-release sensitivity (Appendix C.4, Table 7): Without retraining, we let the opening resistance decrease linearly to zero over a handle displacement DD after release begins (D=0D = 0 is the original model). Three policy seeds were each evaluated on 1,024 episodes with ρ ∈ [1.25, 2.0]; the D=0D = 0 row comes from this separate evaluation, not from Table 1:

    Release distance DDMicrowaveMini fridgeAir fryer
    0 mm (training model)85.3 ± 2.263.2 ± 5.395.7 ± 1.1
    10 mm83.9 ± 1.362.4 ± 4.195.8 ± 1.3
    30 mm80.8 ± 1.757.3 ± 4.595.3 ± 2.3

    Success changes by at most 1.4 points at D=10D = 10 mm and decreases by at most 5.9 points at D=30D = 30 mm.

  • New geometries and automatic initialization (Section 5.3, Table 5; Appendices C.2 and D.1): We added a cabinet with a prismatic drawer and an oven with a revolute door, and a geometry-based algorithm that initializes both arms without manual intervention. With the reward, curriculum, and hyperparameters unchanged, BARS-RL achieves 94.2%/92.7% on the cabinet and 96.6%/92.2% on the oven over three seeds (lower/higher ρ bins). One policy is still trained per object, which we list as a limitation (Appendix B).

(Continued in the next comment.)

AuthorsOpenReview ↗

Response to Reviewer NaGf (3/3)

Q1:

How is each episode initialized? For the holding hand, do you assume that the robot arm is already at a grasping position? Do you assume that the support arm is always applying a downward force at a fixed location with the same hand gesture?

Response:

Initial poses are now selected by an automatic algorithm with common rules and parameters across objects and robots, combining analytical reachability and force-feasibility analysis with geometric checks (Appendix D.1).

  • Task arm: It starts at a grasp derived from the reference grasp of the handle type (Appendix D.1.2), and a scripted warm-up closes the fingers before the policy takes control.
  • Support arm: It starts with four fingers flat on the top surface, without pressing (Appendix D.1.1). The policy then commands both a relative pose and a wrench target (Appendix D.2), so it learns where and how hard to press rather than executing a fixed-location, constant-force press.

Q2:

What explains the failure of BARS-RL without curriculum-learning? How does the learning curves look like? Are baselines not converging, plateaued, or broken? [...]

Response:

We added a termination breakdown of training without the curriculum on the mini fridge (Appendix C.1, Figure 8). Success stays at zero for all plotted seeds, so the policy never receives the success bonus, and each seed fails in a distinct way: most episodes end in excessive body displacement within 200 iterations (seed 6), most time out (seed 5), or an increasing fraction end in simulation divergence (seed 4). In the last case, since every step incurs a time penalty and no success bonus is ever obtained, ending an episode early yields a higher return, and the policy learns to terminate early. Training therefore does not plateau at partial success; it never discovers the coordinated pulling and pressing that the final task distribution requires.

The curriculum separates these skills: anchored opening, then unanchored opening that base friction can support, then support-dependent episodes (Section 4.3). Success thresholds of 70% and 80% gate entry into Phases 1 and 2 (Table 13). With this staged training, BARS-RL reaches the success rates in Table 1.

Q3:

What evidence isolates the contribution of slip-margin reward?

Response:

We removed only cbasec_\text{base}, the base-slip cost corresponding to the slip-margin reward, while retaining cbodyc_\text{body} and all other terms (Section 5.3, Table 4).

MethodMicrowaveMini fridgeAir fryerAvg.
w/o cbasec_\text{base}61.8 ± 8.031.7 ± 7.686.2 ± 2.159.9
BARS-RL85.3 ± 2.068.4 ± 3.396.7 ± 0.683.5

Unlike cbodyc_\text{body}, which responds after the body moves, cbasec_\text{base} penalizes proximity to the slip boundary before displacement accumulates. Including it improves average success by 23.6 points and reduces seed-to-seed variability on every object.

We hope these results address your concerns. If they resolve the questions about novelty and validation, we would be grateful if you would reconsider your assessment. If any concern remains, please let us know and we will address it during the discussion period.

Reviewer_ZdCmOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies bimanual opening of unanchored articulated objects, where one arm opens the moving part while the other stabilizes the object body. It proposes BARS-RL, combining a stabilization reward with a curriculum based on the resistance-to-friction ratio.

Critical strengths

The paper addresses a practical setting where opening an object can move its entire body, making stabilization an essential part of the task. The physical analysis leads to a clear curriculum design, and the comparison with the resistance-based curriculum provides evidence that this choice improves learning.

Critical weaknesses

Main experiments use three simulated objects with a simplified breakaway model and a separate policy for each object, so the results provide limited evidence about broader applicability. Also, the reported ablations mainly examine the curriculum and do not isolate the contribution of the base-slip cost from the other stabilization reward terms.

Questions for authors

Have you tested sensitivity to a gradual or direction-dependent release model, since this would help assess how much the results depend on the discrete breakaway assumption?

4 comments
AuthorsOpenReview ↗

Response to Reviewer ZdCm (1/2)

We thank the reviewer for the constructive feedback. We have conducted additional experiments and made revisions to address the weaknesses identified in your review. Below, we respond to each concern and question. All changes in the revised PDF are marked in red.

W1:

Main experiments use three simulated objects with a simplified breakaway model and a separate policy for each object, so the results provide limited evidence about broader applicability.

Response:

To address broader applicability, we evaluated new object geometries covering both joint types, developed a geometry-based initialization algorithm to eliminate per-object manual pose tuning, and validated simulation-trained policies on real-world appliances.

  • New objects and geometry-based initialization (Section 5.3, Table 5; Appendices C.2 and D.1): We added a cabinet with a prismatic drawer and an oven with a revolute door from ArtVIP to test whether the pipeline could be reused across new geometries for both joint types. To enable this reuse without manual pose tuning, we developed a geometry-based initialization algorithm. Specifically, the algorithm automatically determines both arms’ initial poses through analytical reachability and force-feasibility analysis, together with geometric contact and collision checks, without any manual intervention. Using the same algorithm and parameter values, we then trained policies for the new objects without changing the reward, curriculum, or hyperparameters. Success rates (%) over three seeds:
Objectρ < 1.25ρ ≥ 1.25
Cabinet (prismatic)94.2 ± 1.192.7 ± 1.6
Oven (revolute)96.6 ± 0.492.2 ± 1.2
  • Real-world applicability (Section 5.2, Table 2, Figure 5): To further assess applicability beyond simulation, we deployed the simulation-trained policies on real-world appliances. BARS-RL succeeded in 9 of 10 microwave trials and all 10 air-fryer trials, whereas single-arm RL dragged the body and failed in all 20 trials. Together, these results demonstrate the applicability of BARS-RL even to real-world appliances, where the learned pulling–stabilization coordination remains effective under physical closure and contact dynamics.

  • Reuse across objects: Although BARS-RL requires one policy per object (Appendix B), the initialization algorithm, reward, curriculum, and hyperparameters can all be reused. The new results therefore support the generality of our pipeline across compatible articulated objects.

W2:

Also, the reported ablations mainly examine the curriculum and do not isolate the contribution of the base-slip cost from the other stabilization reward terms.

Response:

To isolate the base-slip cost’s contribution, we added a controlled ablation that removes only cbasec_\text{base} while retaining cbodyc_\text{body} and all other terms (Section 5.3, Table 4).

MethodMicrowaveMini fridgeAir fryerAvg.
w/o cbasec_\text{base}61.8 ± 8.031.7 ± 7.686.2 ± 2.159.9
BARS-RL85.3 ± 2.068.4 ± 3.396.7 ± 0.683.5

While cbodyc_\text{body} responds only after the body moves, cbasec_\text{base} penalizes forces approaching the slip boundary before displacement accumulates. This earlier feedback complements the reactive displacement cost by providing an explicit stabilization signal before body movement. Its practical benefit is reflected in the ablation: including cbasec_\text{base} improves average success by 23.6 percentage points and reduces seed-to-seed variability.

(Continued in the next comment.)

AuthorsOpenReview ↗

Response to Reviewer ZdCm (2/2)

Q1:

Have you tested sensitivity to a gradual or direction-dependent release model, since this would help assess how much the results depend on the discrete breakaway assumption?

Response:

We tested the model against real closure measurements and gradual-release variants, and clarify its directional dependence below.

  • Real closure measurements (Section 5.3, Figure 6): We measured handle-pulling forces on all three real-world appliances and compared them with simulations calibrated to the measured breakaway forces. In both settings, the force built up and then dropped sharply at release. This pattern also appeared in the magnetically latched mini fridge, whose door snapped open despite continuously decaying magnetic attraction. These observations therefore support the model as an approximation of the measured release behavior.

  • Gradual-release sensitivity (Appendix C.4, Table 7): To test whether performance depends on abrupt release, we evaluated the trained policies without retraining under gradual-release variants. After latch release begins, the opening resistance decreases linearly to zero over a handle displacement of DD, measured from the onset of release. Thus, D=0D = 0 recovers our original discrete breakaway model, while D=10D = 10 and 3030 mm retain a gradually decreasing resistance over the corresponding travel distances. For each object and each DD, we used the same object model and initialization conditions and evaluated three policy seeds for 1,024 episodes per seed (3,072 episodes in total). We report mean success rates and standard deviations across seeds for ρ∈[1.25,2.0]\rho \in [1.25, 2.0]. The D=0D = 0 results below are from this separate three-seed evaluation, rather than the five-seed evaluation in Table 1:

Release distance DDMicrowaveMini fridgeAir fryer
0 mm (training model)85.3 ± 2.263.2 ± 5.395.7 ± 1.1
10 mm83.9 ± 1.362.4 ± 4.195.8 ± 1.3
30 mm80.8 ± 1.757.3 ± 4.595.3 ± 2.3

Success changes by at most 1.4 percentage points at D=10D = 10 mm and decreases by at most 5.9 at D=30D = 30 mm. Thus, within the tested range, the policies remain effective under gradual release, with only a moderate decrease at D=30D = 30 mm. The 19 of 20 successful real-world trials provide complementary evidence under physical closure dynamics (Section 5.2).

  • Direction-dependent sensitivity: Our model already accounts for pulling direction through force projection: a pull at angle θ\theta requires Fres/cos⁡θF_\text{res}/\cos\theta to trigger release (Eq. 4, Section 4.1). For small angular misalignment, this requires only a modest increase in force, while the off-axis penalty encourages the policy to maintain alignment with the opening direction (Section 4.2). The practical adequacy of this approximation is further supported by our real-world experiments: simulation-trained policies succeeded in 19 of 20 trials on physical appliances, despite their more complex closure mechanics. These results demonstrate successful deployment under the tested real-world conditions without explicitly modeling the detailed angle-dependent mechanics of each closure.

We hope these additional results address your concerns. If so, we would appreciate it if you would consider updating your score, and we welcome any further questions.

Thanks for the update. The new object and real-world evaluations, controlled base-slip ablation, and gradual-release tests address my main concerns, so I’m raising my rating from 2 to 3.

Thank you for reviewing our response and for raising your rating. Your comments on broader applicability and the base-slip ablation directly led to experiments that strengthened the paper.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Official Comment by Authors

Thank you for the clarification, and for noting that you do not see many issues at the methodological level.

Real-robot trials. Each appliance was placed at the same nominal position for all trials, and the supplementary video shows one representative trial. The object was placed by hand before each trial without precise alignment, so its initial pose differed from trial to trial by up to about 11 mm and 2∘2^\circ (experiment logs). The policy handled this variation with no calibration or adjustment between trials and opened the appliances in 19 of 20 trials (Table 2).

Large-scale benchmark. We agree that it would strengthen the evidence. To our knowledge, existing large-scale articulated-object benchmarks fix the object body to the world and let joints move without resistance once actuated (Section 2), so they do not contain the setting this paper studies. Within the discussion period, we extended the evaluation to five simulated objects, two of them added without any tuning (Table 5), and transferred simulation-trained policies to two real appliances (Table 2). We see a large-scale benchmark of unanchored objects with opening resistance as the natural next step, and the formulation and initialization rules in this paper are meant to make it possible.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.