- Rating
- 3
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
TPR adds a lightweight residual controller on top of a frozen base policy for precise contact-rich manipulation tasks. A progress estimator maps a history of tactile and robot state observations to a scalar interaction process coordinate, trained with weak supervision from relative time within recorded episodes plus endpoint and temporal order auxiliary losses. The coordinate conditions the residual actor, supplies a progress difference term in the PPO reward, and at deployment a hard threshold on it gates whether the bounded residual is applied. Evaluation covers eight ManiFeel tasks, four ManiSkill3 tasks and four real world tasks on an ARX X5 with Xense tactile sensors.
Critical strengths
- The decomposition of progress into three distinct roles, conditioning, reward, and gating, is a clean idea validated by ablation results.
- Evaluation is in thorough in general, with 100 real-world trails per method per task.
Critical weaknesses
- The progress coordinate is defined as relative time within a recorded episode, which presumes a single canonical timeline per task. But when the task has multi-modality, i.e., there are multiple valid strategies to complete the task of differing length, or if demonstrations are collected at different speeds, then the same tactile and state history would map to different targets across episodes, causing ambiguity.
- Real-world success rates are pretty low, and lacks failure mode analysis.
Questions for authors
- Would there be an continuous alternative to the hard gate and what would be some trade-offs, if any?
- Why is g_t=1 during post training but gated at deployment, and why is the evaluation correction gain larger than the training gain?
- Related work on residual correction for contact-rich manipulation should discuss Compliant Residual DAgger: Improving Real-World Contact-Rich Manipulation with Human Corrections (NeurIPS 2025). It is closely related, it also learns a force-aware residual policy over a frozen pretrained base policy for precision contact-rich tasks, and its evaluation tasks also include gear insertion.
2 comments
Responses to Reviewer jfvc (2/2)
Q2-1: "Why is g_t=1 during post training but gated at deployment?"
A2-1: i) Design rationale. We keep during training to learn residual corrections without a hard progress threshold. Corrections during contact directly affect task completion, whereas those during approach may have little effect on return and thus receive a weaker learning signal. Progress gating at deployment suppresses these less-constrained approach-phase corrections, separating residual learning from intervention timing (Lines 221–228, 458–461).
ii) Training and deployment comparison. We compare three training and deployment settings on the four-task ManiFeel subset. Whenever enabled, progress gating uses the same threshold, .
| Training | Deployment | Peg ↑ | USB ↑ | Plug ↑ | Gear ↑ | Average ↑ |
|---|---|---|---|---|---|---|
| Always active | Always active | 45 | 50 | 65 | 59 | 54.8 |
| Progress gate | Progress gate | 50 | 58 | 73 | 69 | 62.5 |
| Always active | Progress gate | 52 | 62 | 76 | 72 | 65.5 |
The two matched settings reduce average success by 10.7% (54.8% vs. 65.5%) and 3.0% (62.5% vs. 65.5%), respectively. These results support learning residual corrections without a hard threshold during training and scheduling their application through progress gating at deployment. We have included this comparison in the revised manuscript (Table 8; Lines 439–444, 458–461).
Q2-2: "why is the evaluation correction gain larger than the training gain?"
A2-2: i) Design rationale. During training, residual actions are sampled from a Gaussian distribution, so a smaller gain limits exploration-induced perturbations to the base actions. At evaluation, a larger gain strengthens deterministic corrections while retaining the same residual budgets and clipping bounds (Eq. 5).
ii) Gain sweep. In our ManiSkill3 experiments, we use a training gain of and an evaluation gain of . To assess sensitivity to the evaluation gain, we evaluate the same trained checkpoint with different gains while keeping the training gain fixed.
| Evaluation gain | Average SR (%) ↑ |
|---|---|
| 0.8 | 45.2 |
| 1.0 | 45.4 |
| 1.2 | 45.9 |
| 1.4 | 45.5 |
| 1.6 | 45.1 |
Average success varies by only 0.8% across the tested gains, indicating low sensitivity to the evaluation gain within this range. We have clarified the training and evaluation gains and added the sweep results to the revised manuscript (Sec. 3.3, Appendix B.2, Appendix C.2 and Table 13, Lines 251–254, 803–805, 868–875).
Q3: "Related work on residual correction for contact-rich manipulation should discuss Compliant Residual Dagger[3]."
A3: Thank you for pointing us to this work. Both methods refine a pretrained policy through residual correction, but differ in: i) Learning: CR-DAgger learns from human corrections; TPR uses PPO with geometric and progress-difference rewards. ii) Control: CR-DAgger predicts pose corrections and target wrenches for compliant control; TPR produces bounded action corrections from tactile feedback and estimated progress. iii) Timing: CR-DAgger learns zero corrections from labeled examples; TPR explicitly gates deployment corrections using progress, which also conditions the policy and shapes its reward. We have cited CR-DAgger and discussed its differences from TPR in Related Work (Lines 082–085).
[1] Iretiayo Akinola et al. TacSL: A Library for Visuotactile Sensor Simulation and Learning. IEEE Transactions on Robotics, 2025.
[2] Kevin Black, Manuel Y. Galliker, and Sergey Levine. Real-Time Execution of Action Chunking Flow Policies. In NeurIPS, 2025.
[3] Xiaomeng Xu, Yifan Hou, Zeyi Liu, and Shuran Song. Compliant Residual DAgger: Improving Real-World Contact-Rich Manipulation with Human Corrections. In NeurIPS, 2025.
Responses to Reviewer jfvc (1/2)
We are grateful to Reviewer jfvc for the thoughtful comments and suggestions. We respond to the comments below and invite the reviewer to consult the revised PDF for the related changes.
W1: "Different execution speeds and valid strategies may assign conflicting temporal labels to similar interaction histories."
A1: Thank you for highlighting this limitation. We use relative-time weak supervision to avoid time-consuming manual progress annotation. We acknowledge that this trade-off can introduce conflicting labels for similar tactile and state histories when demonstrations differ in local execution speed or strategy, potentially reducing estimation accuracy. The estimator therefore depends in part on temporal consistency across demonstrations. We have clarified this limitation in the revised manuscript (Appendix E, Lines 1028–1031).
W2-1: "Real-world success rates remain low..."
A2-1: In real-world tasks, tactile sensing errors, execution delays, and robot motion jitter can reduce task success, with jitter particularly disruptive to fine alignment and contact adjustment. These remain recognized challenges in real-world robotic manipulation [1, 2].
Nevertheless, under the same evaluation protocol, TPR achieves an average success rate of 47.8%, compared with 34.0% for the strongest baseline, an improvement of 13.8% (Table 3).
W2-2: "...failure-mode analysis is missing."
A2-2: Thank you for pointing this out. We have expanded the failure analysis based on the trajectory visualizations in Appendix D.1. These cases show that when the base policy produces large position or orientation errors, the bounded residual correction may be insufficient to resolve them, resulting in persistent misalignment or incomplete insertion or engagement.
Q1: "Would there be a continuous alternative to the hard gate and what would be some trade-offs?"
A1: Thank you for this suggestion. A continuous alternative is a piecewise-linear gate, gt=clip((p^t−τ+δ)/(2δ)0,1), where δ>0 controls the half-width of the transition interval around τ.
This makes correction activation smoother, but allows partial corrections below τ and attenuates corrections just above it, potentially applying correction before it is needed or weakening it when full correction is useful. It also introduces an additional hyperparameter, δ, that requires tuning. We will explore this promising direction in future work.