acceptodds
Desk rejected at ICLR 2027

Pushing Your Robot Policy toward Its Speed Limits via Coupled Gain–Duration Optimization

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Robot policies learned from demonstrations can perform increasingly complex tasks with high reliability, but their execution speed often remains constrained by the demonstrations from which they are learned. We argue that an important bottleneck lies one level below the policy—the low-level controller that determines how quickly and how long each action is physically realized. We introduce Coupled Gain–Duration Optimization (CGDO), which accelerates pretrained policies by jointly optimizing the controller gain and execution duration for each action with reinforcement learning (RL) without modifying the policy. These two variables respectively govern how aggressively the commanded action target is approached and how long the action is executed, and their effects are tightly coupled, making joint optimization critical for pushing execution toward the speed limits without sacrificing task performance. CGDO is policy-agnostic by design and can be combined with other RL methods for further acceleration. Across a suite of experiments, CGDO achieves substantially shorter execution times than other compared methods, while in most cases can also improve success rates.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Robot policy acceleration, Reinforcement learning, Low-level controller

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_BFdSOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the issue of imitation learning policies executing slower than necessary and proposes an approach to accelerate performance at the controller level beneath the policy layer. The core method, CGDO (Coupled Gain–Duration Optimization), freezes the pre-trained base policy and additionally learns an execution adapter that outputs pairs of "controller gain g and execution duration d" for each predicted action block. Training is conducted using off-policy RL (SAC) within a semi-Markov decision process (SMDP) framework, with physical time cost as the reward. Section 3 of the paper also provides an analysis of the motivation. The experiments cover nine simulation tasks and three real-world tasks on the Piper robotic arm; baselines include DemoSpeedup, SpeedAug, SAIL, RACE, SpeedTuning, and policy-level RL methods DSRL/ResFiT. Results show that CGDO achieves 1.42–12.29× speedup compared with the base policy while all task success rates are maintained or improved; when combined with policy-level RL, it achieves the best TPR in eight out of nine tasks.

Critical strengths

i) The problem is identified in a novel and persuasive way, and the method can be combined with previous work. It shifts the acceleration issue from 'what action the policy outputs' to 'how the controller executes the action,' addressing the real pain point of commanded-pose policies where the reached-pose assumption does not hold, while keeping the base policy frozen and allowing for further improvements when combined with DSRL/ResFiT and RTC. ii) The motivation analysis is solid, and the experiments are extensive. Section 3 and the gain-duration coupling analysis in Appendix F clearly explain why joint learning is necessary; the experiments span two types of policy architectures, two categories of action labels, nine simulation tasks, and real hardware, including ablations on multiple seeds variance, time penalty sensitivity, and damping learning, which enhance the credibility of the conclusions.

Critical weaknesses

i) The real-machine verification is insufficient. The real-machine tests involved only 3 tasks, 10–20 trials per method, using a single Piper arm, with no observed improvement in success rate, and the statistical power is insufficient to support strong conclusions at the "deployed system" level. ii) Each task requires online RL retraining, and the costs and generalizability are not sufficiently explained. Each task needs to train the execution policy separately, and combinations such as DSRL on Can also require hyperparameter configuration adjustments; the SR for the three seeds of Relocate fluctuates between 74–88%, indicating limited training stability. The authors acknowledge that cross-task/cross-dynamics transfer has not been validated, but the paper does not report the wall-clock cost of online interactions or a comparison of sample efficiency, making it difficult to evaluate 'the actual overhead of this method relative to retraining the policy itself.'

Questions for authors

i) Please provide the total amount of online interaction and wall-clock time for training the execution strategy of each task, and compare it with the training cost of strategy-level RL methods like DSRL/ResFiT; how sensitive is the adapter training to hyperparameters? ii) On Adroit, RACE directly uses a commanded-pose target for waypoint planning, which leads to infeasibility and all fallback to a 10ms joint-PD controller — should this result be attributed to the method itself or the transfer setup? Did you try retraining the Adroit policy using the reached-pose target originally required by RACE? iii) In real-world experiments, how are safety margins for gain and duration determined? Under high gain, did you ever encounter joint torque saturation, vibration, or hardware protective shutdown?

5 comments
AuthorsOpenReview ↗

Response to Reviewer BFdS (1)

We appreciate your insightful feedback. We have carefully considered your comments and provided a point-by-point response, and we hope it addresses your concerns.

Q1. Can you evaluate more real-world tasks with more trials per method?

The real-machine verification is insufficient. The real-machine tests involved only 3 tasks, 10–20 trials per method, using a single Piper arm, with no observed improvement in success rate, and the statistical power is insufficient to support strong conclusions at the "deployed system" level.

We add a real-world task requiring precise cube alignment, with 50 trials per method. CGDO reduces mean successful wall-clock completion time from 9.36s to 1.32s (7.11× speedup), while observed success increases from 82% (41/50) to 88% (44/50).

MethodSuccess Rate (%) ↑ATR (s) ↓
Fixed82% (41/50)9.36 ± 1.08
CGDO88% (44/50)1.32 ± 0.06

( ATR: mean successful physical completion time )

Q2. What is CGDO’s interaction overhead, and how sensitive is training to hyperparameters?

Each task requires online RL retraining, and the costs and generalizability are not sufficiently explained. Each task needs to train the execution policy separately, and combinations such as DSRL on Can also require hyperparameter configuration adjustments; the SR for the three seeds of Relocate fluctuates between 74–88%, indicating limited training stability. The authors acknowledge that cross-task/cross-dynamics transfer has not been validated, but the paper does not report the wall-clock cost of online interactions or a comparison of sample efficiency, making it difficult to evaluate 'the actual overhead of this method relative to retraining the policy itself.'

Please provide the total amount of online interaction and wall-clock time for training the execution strategy of each task, and compare it with the training cost of strategy-level RL methods like DSRL/ResFiT; how sensitive is the adapter training to hyperparameters?

Figure 6b (Appendix C.1) compares sample efficiency on Lift. CGDO reaches a 200-episode average return of −40 after approximately 21.6k online environment steps, compared to 832k for DSRL, a 38.5× reduction during the online phase. Even including initial warm-up data collection, CGDO requires about 103.9k steps, roughly 1/8 of DSRL’s online interactions alone.

Notably, DSRL is the strongest RL baseline in this comparison, achieving higher returns and sample efficiency than both ResFiT and SpeedTuning.

In addition, we use a single shared set of hyperparameters within each benchmark, supporting easy reuse across tasks (Appendix A.1).

Q3. Does the Adroit RACE result reflect the method or the transfer setup?

On Adroit, RACE directly uses a commanded-pose target for waypoint planning, which leads to infeasibility and all fallback to a 10ms joint-PD controller — should this result be attributed to the method itself or the transfer setup? Did you try retraining the Adroit policy using the reached-pose target originally required by RACE?

RACE is designed for policies trained on reached targets. Directly applying it to commanded targets can alter the intended motion, because these commands need not correspond to states attained during demonstration. When planning fails, RACE defaults to PD control, according to the paper [1]. Therefore, RACE is not directly applicable to commanded targets.

In our manuscript, we use reached targets on RoboMimic and MimicGen, and commanded targets on Adroit. The former use RACE’s intended target type, whereas the latter do not. On Adroit, nearly all RACE planning attempts fail, triggering PD fallback as specified in its original paper, so the low scores are not surprising.

We will make this distinction explicit in the revised manuscript.

That said, to directly address this concern, we follow the reviewer's suggestion and re-train the Adroit Door base policy using reached targets. Results show CGDO still outperforms RACE in both success rate and speed, consistent with our findings on RoboMimic and MimicGen.

MethodSuccess Rate (%) ↑ATR (s) ↓Throughput (TP₆₀₀) ↑
RACE73.22 ± 2.55%0.855 ± 0.007378.55±22.83
CGDO94.89 ± 2.50%0.670 ± 0.010774.29 ± 61.74

( TP₆₀₀: estimated successful completions per 600 seconds )

[1] Time Optimal Execution of Action Chunk Policies Beyond Demonstration Speed. ICLR 2026.

AuthorsOpenReview ↗

Response to Reviewer BFdS (2)

Q4. How are controller bounds and safety margins determined?

In real-world experiments, how are safety margins for gain and duration determined? Under high gain, did you ever encounter joint torque saturation, vibration, or hardware protective shutdown?

All six joints use a proportional gain of Kp=10K_p=10. CGDO selects a shared proportional gain within Kp∈[8,12]K_p\in[8,12] and action execution durations of 1–50 ms.

We select the gain bounds empirically through parameter sweeps. At larger gains, we observe joint vibrations at approximately 35 Hz. The selected range allows gain adaptation while limiting vibration.

We use a manual stop procedure when vibration is observed. No such stop is triggered within the selected gain range, where we observe no signs of torque saturation and no hardware protective shutdowns.

AuthorsOpenReview ↗

Gentle Reminder

Thanks for your time and effort in reviewing our work. As the discussion draws to a close, we would be grateful if you could have a look at the rebuttal and let us know if any concerns remain.

Following our point-by-point response, we would like to share a new real-world evaluation on clothes folding, a long-horizon, contact-rich task involving deformable-object manipulation. These task properties makes clothes folding a particularly demanding test of CGDO, further broadening the scope of our real-world evaluation, and better addressing the concern you raised in weakness (i).

MethodSuccess Rate (%) ↑ATR (s) ↓Throughput (TP600TP_{600}) ↑
Fixed98% (49/50)37.45 ± 1.6215.73 ± 0.31
CGDO94% (47/50)8.94 ± 1.8460.84 ± 3.76

( ATR: mean successful physical completion time. TP600TP_{600}: estimated successful completions per 600 seconds )

With π0.5\pi_{0.5} as the base policy and 50 evaluation trials per method, CGDO reduces mean successful completion time from 37.45s to 8.94s (4.19× speedup) and increases throughput from 15.73 to 60.84 (3.87×). These results further support CGDO’s practical value for challenging tasks in the real world. We also provide an experiment video on this anonymous supplementary website and would appreciate your taking a look.

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

This paper addresses the issue where the execution speed of robot policies learned via imitation is constrained by the speed of the demonstrations. The authors propose "Coupled Gain–Duration Optimization" (CGDO), which keeps a pre-trained policy fixed while using reinforcement learning to optimize a low-level controller; policy acceleration is achieved by incorporating a physical-time objective within a semi-MDP framework. The approach is evaluated across various simulated tasks and real-world robotic manipulation scenarios.

Critical strengths

(i) The research problem holds practical significance. By re-examining the acceleration of robot policies from the perspective of low-level controllers, the authors offer a complementary optimization approach to existing policy-level acceleration methods.

(ii) The method design is clear. CGDO optimizes the execution process without requiring modifications to the pre-training strategy, and it models variable execution durations through a semi-MDP framework. This approach offers a degree of modularity, as it can be integrated with methods such as DSRL, ResFiT, and RTC.

(iii) The experimental coverage is comprehensive. The authors conducted evaluations on benchmarks such as RoboMimic, MimicGen, and Adroit, providing thorough experimental evidence for the effectiveness of their method.TeX is supported

Critical weaknesses

(i) The methodological novelty is limited. CGDO primarily integrates existing techniques in controller parameter optimization, adaptive action duration, and semi-MDP RL without introducing a distinctly new algorithmic mechanism. A clearer characterization of the unique algorithmic contributions beyond combining existing techniques would strengthen the paper.

(ii) As the core motivation of the paper, the incremental value and general applicability of joint optimization remain unclear. Figure 3 indicates that the majority of acceleration gains stem from "duration-only optimization," and the relevant ablation studies focus primarily on the "Lift" task. It is recommended that the authors extend their evaluation to a broader range of tasks involving diverse contact conditions to verify the independent contribution and general effectiveness of the joint optimization approach.

(iii) The baseline comparisons are incomplete and raise fairness concerns. A direct comparison with the closely related Speedup Patch is missing, while RACE falls back to PD control throughout the Adroit evaluation. More rigorous comparisons with closely related methods under consistent and appropriate execution conditions are needed to substantiate the reported advantages.

(iv) There is a lack of thorough analysis regarding the trade-off between execution speed and task reliability. Specifically, the CGDO approach lacks explicit constraints on success rates. In some experiments, the integration of CGDO leads to increased execution speed but a decline in success rate, for instance, as observed with DSRL+CGDO on the "Square" and "Stack" tasks. The authors are advised to further analyze this speed–success trade-off and report performance under varying reliability requirements.

(v) The generalization capability and real-world validation remain limited. CGDO relies on task-specific training, with no evaluation of transfer across tasks or robot platforms. Real-world experiments involve only three tasks on a single platform, with 10–20 trials per method. Further evaluations of cross-task transfer and robustness under challenging execution conditions would provide stronger evidence of practical applicability.

Questions for authors

See Weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer RGhp (1)

We thank the reviewer for the constructive feedback.

Q1. What is CGDO’s technical contribution beyond combining existing techniques?

The methodological novelty is limited. CGDO primarily integrates existing techniques in controller parameter optimization, adaptive action duration, and semi-MDP RL without introducing a distinctly new algorithmic mechanism. A clearer characterization of the unique algorithmic contributions beyond combining existing techniques would strengthen the paper.

CGDO’s core novelty lies in showing why these existing components must be combined to address robot execution speedup. Not an arbitrary assembly, their integration follows the requirements of the problem.

For example:

  • Our goal is to accelerate execution without distorting motions, so we jointly optimize controller gain and duration (as motivated in Section 3).
  • We aim to support both commanded- and reached-pose targets, so we learn from task reward using RL.
  • Action durations vary across steps, so we formulate a semi-MDP with physical-time rewards and discounting.
  • Robot rollouts are costly, so we adopt off-policy SAC to reuse collected experience.

Sometimes, method novelty does not always require inventing a new primitive. For instance, InstructGPT [1] combines supervised fine-tuning, reward modeling, and PPO, yet this synthesis made RLHF as a cornerstone of modern LLM post-training. Similarly, DigiRL [2] integrates advantage-weighted RL, autonomous evaluation, and curriculum learning, marking a watershed transition for agents from imitation to autonomous online RL.

Similarly, CGDO identifies the gain–duration interface as a key for execution speed and formulates the exact RL pipeline to optimize it.

[1] Training language models to follow instructions with human feedback. OpenAI, NeurIPS 2022.

[2] DigiRL: Training In-The-Wild Device-Control Agents with Autonomous Reinforcement Learning. UC Berkeley, NeurIPS 2024.

Q2. Does joint optimization provide additional benefits beyond duration-only optimization on contact-rich tasks?

As the core motivation of the paper, the incremental value and general applicability of joint optimization remain unclear. Figure 3 indicates that the majority of acceleration gains stem from "duration-only optimization," and the relevant ablation studies focus primarily on the "Lift" task. It is recommended that the authors extend their evaluation to a broader range of tasks involving diverse contact conditions to verify the independent contribution and general effectiveness of the joint optimization approach.

We add an ablation on Adroit Door, a contact-rich task requiring a dexterous hand to unlatch and open a door. All results are evaluated on 300 held-out episodes.

MethodSuccess Rate (%) ↑ATR (s) ↓Throughput (TP₆₀₀) ↑
Fixed (best hyper-params)43.0 ± 5.60.837 ± 0.043172.0 ± 30.0
Gain-only70.0 ± 9.00.851 ± 0.093351.3 ± 78.2
Duration-only92.3 ± 3.00.704 ± 0.023689.9 ± 58.9
Joint CGDO98.0 ± 1.60.519 ± 0.0261070.9 ± 80.3

( ATR: mean successful physical completion time TP₆₀₀: estimated successful completions per 600 seconds )

All learned variants improve success. Unlike on Lift, duration-only accounts for only 42% of CGDO’s ATR reduction on this dexterous task. Joint CGDO achieves both the highest success rate and the lowest ATR among all evaluated variants.

AuthorsOpenReview ↗

Response to Reviewer RGhp (2)

Q3. How do the additional comparisons address the missing SuP baseline and concerns about RACE?

The baseline comparisons are incomplete and raise fairness concerns. A direct comparison with the closely related Speedup Patch is missing, while RACE falls back to PD control throughout the Adroit evaluation. More rigorous comparisons with closely related methods under consistent and appropriate execution conditions are needed to substantiate the reported advantages.

We include a direct comparison with Speedup Patch (SuP) on four tasks. CGDO achieves higher success and throughput on all four tasks.

TaskSR–SuP (%) ↑SR–CGDO (%) ↑TP₆₀₀–SuP ↑TP₆₀₀–CGDO ↑
Lift93.8910084.7959.93
Can76.6798.3347.31284.12
Square39.1173.6714.3243.47
Mug48.8996.6713.94123.47

..., while RACE falls back to PD control throughout the Adroit evaluation.

RACE is designed for policies trained on reached targets. Directly applying it to commanded targets can alter the intended motion, because these commands need not correspond to states attained during demonstration. When planning fails, RACE defaults to PD control, according to the paper [1]. Therefore, RACE is not directly applicable to commanded targets.

In our manuscript, we use reached targets on RoboMimic and MimicGen, and commanded targets on Adroit. The former use RACE’s intended target type, whereas the latter do not. On Adroit, nearly all RACE planning attempts fail, triggering PD fallback as specified in its original paper, so the low scores are not surprising.

We will make this distinction explicit in the revised manuscript.

That said, to directly address this concern, we follow the reviewer's suggestion and re-train the Adroit Door base policy using reached targets. Results show CGDO still outperforms RACE in both success rate and speed, consistent with our findings on RoboMimic and MimicGen.

MethodSuccess Rate (%) ↑ATR (s) ↓Throughput (TP₆₀₀) ↑
RACE73.22 ± 2.55%0.855 ± 0.007378.55±22.83
CGDO94.89 ± 2.50%0.670 ± 0.010774.29 ± 61.74

[1] Time Optimal Execution of Action Chunk Policies Beyond Demonstration Speed. ICLR 2026.

Q4. How does CGDO trade execution speed against task reliability?

There is a lack of thorough analysis regarding the trade-off between execution speed and task reliability. Specifically, the CGDO approach lacks explicit constraints on success rates. In some experiments, the integration of CGDO leads to increased execution speed but a decline in success rate, for instance, as observed with DSRL+CGDO on the "Square" and "Stack" tasks. The authors are advised to further analyze this speed–success trade-off and report performance under varying reliability requirements.

The reward balances speed and reliability. Specifically, with timeout hh, the expected return JJ is J=−c[ptˉ+(1−p)h]J = -c[p\bar{t} + (1 - p)h ], where pp is the success rate and tˉ\bar{t} is the mean successful completion time. Because every failure incurs the full timeout cost hh, a large reduction in tˉ\bar{t} can initially outweigh a modest drop in pp, whereas excessive failures quickly negate the acceleration benefits.

The Lift curves (Appendix C.1, Figure 6a) reflect this. Early in training, faster execution improves return enough to outweigh a temporary success-rate dip. Later, as speedups saturate, improving reliability to avoid the timeout penalty becomes the more rewarding choice, so success recovers toward 100%. Despite a few exceptions (e.g., DSRL+CGDO on Square, Stack), CGDO improves both speed and success across most tasks. (table 1-2 in the manuscript)

For tasks with strict reliability requirements, we find that gradually expanding the controller bounds, rather than applying the full bounds from the start, effectively mitigates early success-rate drops. (see Appendix C.1, Figure 7)

AuthorsOpenReview ↗

Response to Reviewer RGhp (3)

Q5. Can one CGDO controller generalize across tasks?

The generalization capability and real-world validation remain limited. CGDO relies on task-specific training, with no evaluation of transfer across tasks or robot platforms. Real-world experiments involve only three tasks on a single platform, with 10–20 trials per method. Further evaluations of cross-task transfer and robustness under challenging execution conditions would provide stronger evidence of practical applicability.

To assess CGDO’s cross-task generalization, we train a single shared CGDO controller on eight LIBERO-Goal tasks with a frozen π₀.₅. We then evaluate the same controller on two out-of-disturbution tasks excluded from CGDO training, without further adaptation.

Held-out TaskSuccess Rate-Base (%)Success Rate-CGDO (%)Throughput Speedup
task 1: bowl on plate100%96%1.35×
task 2: bottle on rack96%94%1.86×
Combined98%95%1.64×

CGDO achieves 1.64× combined throughput on tasks unseen during CGDO training, with a 3% decrease in success rate, a modest cost for cross-task generalization. These results show that acceleration can transfer to new tasks without task-specific CGDO training.


We also add a real-world task requiring precise alignment, using a π₀.₅ policy and RTC. We evaluate each method over 50 trials. CGDO reduces mean successful wall-clock completion time from 9.36s to 1.32s (7.11× speedup), while observed success increases from 82% to 88%.

MethodSuccess Rate (%) ↑ATR (s) ↓
Fixed82% (41/50)9.36 ± 1.08
CGDO88% (44/50)1.32 ± 0.06
AuthorsOpenReview ↗

Gentle Reminder

Thanks for your time and effort in reviewing our work. As the discussion draws to a close, we would be grateful if you could have a look at the rebuttal and let us know if any concerns remain.

Following our point-by-point response, we would like to share a new real-world evaluation on clothes folding, a long-horizon, contact-rich task involving deformable-object manipulation. These task properties makes clothes folding a particularly demanding test of CGDO, further broadening the scope of our real-world evaluation.

MethodSuccess Rate (%) ↑ATR (s) ↓Throughput (TP600TP_{600}) ↑
Fixed98% (49/50)37.45 ± 1.6215.73 ± 0.31
CGDO94% (47/50)8.94 ± 1.8460.84 ± 3.76

( ATR: mean successful physical completion time. TP600TP_{600}: estimated successful completions per 600 seconds )

With π0.5\pi_{0.5} as the base policy and 50 evaluation trials per method, CGDO reduces mean successful completion time from 37.45s to 8.94s (4.19× speedup) and increases throughput from 15.73 to 60.84 (3.87×). These results further support CGDO’s practical value for challenging tasks in the real world. We also provide an experiment video on the anonymous supplementary website and would appreciate your taking a look.

Reviewer_Dgj2OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CGDO, a method to speed up pretrained robot policies without changing the policies themselves. The main idea is to use RL to jointly learn the controller gain and how long each action should run. Experiments on both simulated and real robots show clear improvements in speed, while keeping good success rates.

Critical strengths
  • Figure 2 gives a concrete reason to consider gain and duration together. Changing either alone alters the path; coordinating them approximately preserves it in free motion, but the same rule breaks during contact. This makes a better case for learning than simply showing that higher gains produce faster movement.

  • The real-robot timings are worth highlighting. On Nest Cups, completion time falls from 12.86s to 4.49s at the same observed 80% success rate, and these measurements include inference and communication.

Critical weaknesses
  • I am not fully convinced that learning gain and duration together is necessary. On Lift, learning duration alone already gives most of the speedup (0.83s vs. 0.63s). It would be helpful to see this comparison on more tasks, especially those involving contact.

  • The simulation uses torque limits that are 20 times higher than the original limits. I wonder how much this affects the results, and whether similar speedups can be achieved with normal limits.

Questions for authors

Please see the weaknesses above. Appendix D.2 specifies durations of 1–50 ms for the real robot, but I could not find the corresponding gain range. What gains were allowed on the Piper arm, and how were those bounds chosen?

3 comments
AuthorsOpenReview ↗

Response to Reviewer Dgj2

We thank the reviewer for the constructive feedback.

Q1. Does joint gain–duration optimization provide additional benefits on contact-rich tasks?

I am not fully convinced that learning gain and duration together is necessary. On Lift, learning duration alone already gives most of the speedup (0.83s vs. 0.63s). It would be helpful to see this comparison on more tasks, especially those involving contact.

We add an ablation on Adroit Door, a contact-rich task requiring a dexterous hand to unlatch and open a door. All results are evaluated on 300 held-out episodes.

MethodSuccess Rate (%) ↑ATR (s) ↓Throughput (TP₆₀₀) ↑
Fixed (best hyper-params)43.0 ± 5.60.837 ± 0.043172.0 ± 30.0
Gain-only70.0 ± 9.00.851 ± 0.093351.3 ± 78.2
Duration-only92.3 ± 3.00.704 ± 0.023689.9 ± 58.9
Joint CGDO98.0 ± 1.60.519 ± 0.0261070.9 ± 80.3

( ATR: mean successful physical completion time TP₆₀₀: estimated successful completions per 600 seconds )

All learned variants improve success. Unlike on Lift, duration-only accounts for only 42% of CGDO’s ATR reduction on this dexterous task. Joint CGDO achieves both the highest success rate and the lowest ATR among all evaluated variants.

Q2. How much do the 20× torque limits affect performance, and does CGDO also accelerate execution under native limits?

The simulation uses torque limits that are 20 times higher than the original limits. I wonder how much this affects the results, and whether similar speedups can be achieved with normal limits.

We evaluate CGDO under both native (1×) and 20× torque limits on Lift, Can, and Square, using 300 episodes per setting.

TaskMethodSuccess rate (%) ↑ATR (s) ↓
Liftbase policy85.77.68
CGDO, 1x limits1000.794
CGDO, 20× limits1000.625
Canbase policy5911.99
CGDO, 1x limits952.506
CGDO, 20× limits98.331.858
Squarebase policy1816.07
CGDO, 1x limits62.677.309
CGDO, 20× limits73.676.654

We relax torque limits in simulation for consistency with baseline implementations that also relax them. Although higher limits improve performance, CGDO under native limits (1×) still achieves 9.67×, 4.78×, and 2.20× speedups over the base policy on Lift, Can, and Square, respectively. In our real-world experiments, torque limits cannot be adjusted because of hardware safety and performance constraints.

Q3. What gain range does CGDO use on the Piper arm, and how are the bounds chosen?

Appendix D.2 specifies durations of 1–50 ms for the real robot, but I could not find the corresponding gain range. What gains were allowed on the Piper arm, and how were those bounds chosen?

All six joints use a proportional gain of Kp=10K_p=10, and CGDO selects a shared gain within Kp∈[8,12]K_p\in[8,12]. We choose these bounds empirically through parameter sweeps. Larger gains produce joint vibration, so the selected range allows gain adaptation while limiting vibration.

Thank the authors for the detailed explanations in addressing my concerns, which have resolved most of my points.

Dear Reviewer Dgj2,

Thank you very much for taking the time to carefully review our responses and for your follow-up. We sincerely appreciate your constructive feedback and are pleased to hear that our clarifications have addressed most of your concerns.

In light of these clarifications, we would be very grateful if you could consider revisiting your overall rating. Thank you again for your time and consideration.

Best regards,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Sincere Thanks and Apologies

We sincerely thank all reviewers and the Area Chair for your dedication, especially amid the surge in submissions this year. We deeply appreciate your constructive feedback, which has been instrumental in helping us improve this work.

Due to hardware resource constraints and a heavy workload, our rebuttal is arriving a bit late, and we hope this does not cause you any inconvenience.

If any concerns remain after reading our responses, please feel free to let us know. We will remain actively online to address them.

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

Related papers

Discussion (0)

Sign in to comment.