acceptodds
Under review as a conference paper at ICLR 2027

Blind Precise Assembly: Proprioceptive Sim-to-Real Peg-in-Hole for Diverse Parts

Abstract

Every robot already carries high-resolution contact sensors: its own joints. Legged robots and humanoids exploit them to walk, run, and track whole-body motions without vision, but there the contacts are large and forceful. Is joint sensing also enough enough for precise, contact-rich assembly, where the robot must find a millimeter-scale opening through light contact? We study this question by training a blind insertion policy that observes only a short history of joint positions, velocities, and its own actions, without visual, force, torque, or tactile sensing. The poses of both the hole and the peg are unknown to the policy, and the grasp is randomized substantially. The policy must therefore find the opening through contact, sensed exclusively through its own proprioception. We train this behavior with reinforcement learning entirely in simulation over randomized parts, grasps, and contacts, and it transfers zero-shot to a real robot and inserts pegs of different shapes and sizes in 74% of trials. Controlled comparisons with matched visual policies in simulation show that although blind policies are slower, they generalize far better to unseen part shapes and sizes. Training such policies needs thousands of different assemblies, each simulated with its millimeter clearance intact, so we also release an open-source generator that builds them procedurally from exact convex pieces for MuJoCo; it is more accurate and faster than convex decomposition or MuJoCo's signed-distance collision.

Primary area:
applications to robotics, autonomy, planning
Keywords:
robot learning, reinforcement learning, sim-to-real transfer, proprioception, contact-rich manipulation, robotic assembly

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_T8wuOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies local, precise peg-in-hole insertion under grasp uncertainty using proprioception alone. The policy uses a temporal convolutional encoder to process a 0.6-second history of joint states, gripper states computed through forward kinematics, and previous actions, and outputs joint-position increments. Training uses PPO with an asymmetric privileged critic and auxiliary contact-state supervision. At deployment, the policy receives no images, external force/torque or tactile measurements, hole pose, grasp pose, or part model. The authors also introduce mjfit, a procedural generator of convex collision geometry that preserves assembly clearances in MuJoCo.

The evaluation includes in-distribution performance, generalization to unseen geometries, analysis of the history representation, and real-world deployment of a frozen policy. In simulation, the proprioceptive and visual policies achieve success rates of 99.0% and 99.5%, with median insertion times of 1.05 s and 0.36 s, respectively. Across five real-world parts, the deployed policy succeeds in 112/151 trials, approximately 74.2%.

The task's scope is important: trials begin with the peg already touching the fixture near the hole. Furthermore, the modality-comparison policies and the deployed policy use different training randomization ranges. The paper therefore demonstrates a local insertion skill under restricted sensing, rather than global hole localization from arbitrary initial configurations.

Critical strengths
  1. The real-world demonstration under restricted sensing is a substantive contribution. The paper provides concrete evidence that a policy can use its own motion feedback to search for and insert into an opening with millimeter-scale clearance, without directly observing the hole pose, grasp pose, or external contact measurements. The real-world evaluation covers four stadium-shaped parts of different sizes and proportions, an unseen circular cross-section, and uncertainty in grasp position and tilt. The 112 successes over 151 trials make this more convincing than isolated demonstrations with a fixed part and grasp. 
  2. mjfit is an independently valuable simulation contribution supported by targeted validation. It addresses the problem that collision-geometry approximation errors can change the outcome of tight assembly. It constructs convex geometry that preserves the intended clearance rather than relying on generic approximate decomposition. Section 2.2 and Appendix B evaluate it using passive probes with analytic answers and report CPU/GPU speed–accuracy trade-offs. The benchmark holds solver settings fixed, accounts for the method's own curve-tessellation error, and explains how solver compliance limits the resolution of some probes.
Critical weaknesses
  1. The modality comparison does not sufficiently control training-time information, limiting the attribution of the generalization advantage. Section 2.3 states that the proprioceptive policy uses a privileged critic with access to relative peg–hole pose and true contact labels, together with auxiliary contact supervision for its history representation. In contrast, the visual policy's critic observes only images, and the manuscript does not clearly establish whether the visual representation receives equivalent auxiliary supervision. Thus, although the task, resets, randomization, rewards, and actions are matched, the training-time information is not. The experiments support the narrower conclusion that the proposed proprioceptive policy outperforms the implemented visual baseline on the tested geometry shifts; they do not isolate sensing modality as the cause. Similar in-distribution success does not rule out effects of the training setup on out-of-distribution generalization. 
  2. The task-level benefits of the key learning components are insufficiently established. The representation analysis is useful: the history latent achieves 70% balanced accuracy on seven insertion modes, compared with 57% from the current observation, and the latent-swap experiment shows that the representation affects actions. However, these analyses do not directly quantify how history and contact supervision improve task success, learning efficiency, or transfer.
Questions for authors
  1. Why does the proprioceptive policy use a privileged critic while the visual policy uses an image-only critic? Does the visual representation receive equivalent contact-state supervision? With training-time information matched and training budgets specified, does the generalization gap on extreme sizes and random convex cross-sections remain? 
  2. Appendix A specifies a 5 s simulation episode limit, whereas Figure 1 includes a successful real-world trajectory with a final frame at 5.5 s. What are the real-world timing origin, timeout threshold, and operational definition of "seated"? 
  3. Section 2.3 states that armature, damping, and friction were identified from real-world measurements, while Section 3.3 states "no real-world data or fine-tuning." Were real-world data used only for dynamics identification, or also for checkpoint selection, controller adjustment, or other decisions?
2 comments
  1. “The modality comparison does not sufficiently control training-time information, limiting the attribution of the generalization advantage. (...) Why does the proprioceptive policy use a privileged critic while the visual policy uses an image-only critic? ”

We agree that the “image-only” critic is a potential confounder, and have re-trained the camera-based policy, giving the critic access to the exact same privileged information as the proprioceptive policy (now, the two policies differ only in the actor’s observations) and updated Figs. 7 and 9, as well as Tab. A4, accordingly. The throughput (Fig. 7) and robustness (Fig. 9) trends both remains similar with the privileged critic. We have also clarified in the manuscript that the camera-based policy is indeed trained with the same contact-supervision loss as the proprioception policy.


  1. “The task-level benefits of the key learning components are insufficiently established. (...) Accompanied analyses do not directly quantify how history and contact supervision improve task success, learning efficiency, or transfer.”

To address this concern, we run two sets of ablation experiments (1) removing information from the observation with the same architecture, and (2) removing components from the architecture (contact supervision, temporal encoder) with the same observation.

  • In terms of removing information from the observation, we find that the biggest loss in performance comes from removing previous action and velocity observations, which deprive the policy of any information which can be used to infer the force. We also verify that the policy is not able to learn a purely open-loop solution (i.e. a “better spiral search”) when given only a clock as input. We added the experimental details in Appendix D1.
Actor inputBest success (%)
Proprio, 30-frame history (Ours)93.3
Proprio, 1 frame67.2
Joint + TCP position, 1 frame (no force information)7.2
Clock only (open loop)3.8
  • In terms of removing components from the architecture, both the contact supervision and temporal encoder play a large role in effectively compressing the history into an effective representation for the policy to use. We added the experimental details in Appendix D2.
PolicyBest success (%)
Encoder + supervision93.3
Encoder, no supervision42.0
No encoder (MLP), no supervision17.2

  1. “Appendix A specifies a 5 s simulation episode limit, whereas Figure 1 includes a successful real-world trajectory with a final frame at 5.5 s. What are the real-world timing origin, timeout threshold, and operational definition of "seated"? This is not necessarily a contradiction, but the protocol needs clarification to interpret the reported real-world success rate.”

We see this task as an indefinite horizon RL problem — the 5s timeout during training is there to speed up convergence. During evaluation, we allow the policy to run until success or termination failure (dropping the peg, or traveling further than a threshold distance away from the initialization) both in simulation and the real world. In the real world, a peg is seated if the entire bottom face of the peg is flush against the entire bottom face of the hole (this is easy to detect visually). Any rollout where the peg becomes jammed and remains only partially inserted is considered a failure.


  1. “Section 2.3 states that armature, damping, and friction were identified from real-world measurements, while Section 3.3 states "no real-world data or fine-tuning." Were real-world data used only for dynamics identification, or also for checkpoint selection, controller adjustment, or other decisions? ”

We have revised the language in Section 3.3 to say “without fine-tuning the policy on any real-world data”. Dynamics identification is the component of our system that requires real-world data (position response of the robot arm moving in free space). For checkpoint selection, we deployed the best checkpoint (measured by training time success rate in sim).

I acknowledge the author's re matching of visual strategy training information and the addition of task level ablation with historical input and contact supervision. I maintain a score of 3 (weak acceptance) and request the author to clarify in the final draft the stopping rules for evaluation failure due to lag, as well as the training budget and statistical criteria for the new experiments.

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

The paper presents a reinforcement learning approach for precise peg-in-hole assembly using only robot proprioception. Its main contributions are a contact-aware policy trained entirely in simulation with auxiliary contact supervision and mjfit, a procedural geometry generator enabling accurate, efficient simulation of diverse tight-clearance assemblies. Validation includes simulation comparisons against visual policies and scripted search, and zero-shot real-world deployment achieving 74% success across 151 trials involving five peg geometries.

Critical strengths
  1. Comprehensive evaluation and real-world validation. The paper provides extensive simulation experiments comparing the proposed proprioceptive policy with visual and scripted baselines, including generalization tests on unseen geometries and analyses of learned behaviors. The zero-shot real-world deployment further demonstrates the practical feasibility of the approach, strengthening the credibility of the findings.

  2. Broader utility of mjfit. The proposed mjfit provides an efficient and accurate framework for generating diverse assembly geometries with precise collision modeling. Beyond supporting the proposed method, this tool has the potential to benefit broader research on contact-rich manipulation and sim-to-real robot learning.

Critical weaknesses
  1. Overstated novelty and limited methodological contribution. The paper's novelty claims appear overstated, given existing research on proprioception-based peg-in-hole insertion [1-5]. The proposed methodology largely follows conventional reinforcement learning practices, including PPO, asymmetric actor-critic training, and auxiliary supervision, without introducing substantial algorithmic innovations. This raises concerns about the paper's novelty and its contribution beyond existing approaches.

  2. Limited baselines and insufficient in-depth analysis. The experimental comparisons are limited to a relatively simple wrist-camera CNN policy and scripted search, without considering stronger visual baselines (e.g., VLM/VLA-based methods), alternative camera viewpoints, or force-based approaches. Furthermore, all baseline comparisons are conducted in simulation, with no real-world comparative evaluation. While the paper demonstrates promising results, it provides limited mechanistic insight into why proprioception alone is sufficient for precise assembly and what drives its superior generalization. These limitations weaken the evidence supporting the paper's central claims.

[1] Pan J, Coros S, Malik J, et al. Beyond Binary: Sim-to-Real Dexterous Manipulation with Physics-Grounded Contact Representation[J]. arXiv preprint arXiv:2605.28812, 2026. [2] Shan S, Pham Q C. Fine robotic manipulation without force/torque sensor[J]. IEEE Robotics and Automation Letters, 2023, 9(2): 1206-1213. [3] Wang C, Luo H, Zhang K, et al. Pomdp-guided active force-based search for robotic insertion[C]//2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2023: 10668-10675. [4] Zhang, Tie, Xiaohong Liang, and Yanbiao Zou. "Robot peg-in-hole assembly based on contact force estimation compensated by convolutional neural network." Control Engineering Practice 120 (2022): 105012. [5] Park, Hyeonjun, et al. "Compliance-based robotic peg-in-hole assembly strategy without force feedback." IEEE Transactions on Industrial Electronics 64.8 (2017): 6299-6309.

Questions for authors
  1. Could the authors clarify the key technical distinctions between the proposed approach and existing proprioception-based peg-in-hole methods? In particular, what constitutes the main methodological novelty beyond conventional RL training with auxiliary supervision?

  2. The proprioceptive policy benefits from a privileged critic and auxiliary contact supervision, whereas the visual policy uses an image-only critic. How can the authors disentangle the effect of sensing modality from differences in training supervision when attributing superior generalization to proprioception?

  3. Could the authors provide comparisons with stronger visual or force-based insertion policies, ideally including real-world evaluations? Such evidence would be important for establishing whether the reported advantages extend beyond the relatively simple simulation baselines.

  4. Could the authors provide further ablations, particularly regarding contact supervision and observation history, to explain what enables the policy to infer contact states and generalize to unseen geometries? This would strengthen the central claim that proprioception provides sufficient information for precise assembly.

6 comments

We thank the reviewer for giving such constructive feedback. We address the concerns below:


  1. “Overstated novelty and limited methodological contribution. (...) Could the authors clarify the key technical distinctions between the proposed approach and existing proprioception-based peg-in-hole methods? In particular, what constitutes the main methodological novelty beyond conventional RL training with auxiliary supervision?”

We do not claim algorithmic novelty. Our contribution is a system showing that proprioception alone is enough for general, precise sim-to-real insertion. It differs from prior systems in two ways at once. It uses less information: no tactile, force/torque, or visual sensing, and no hole, part, or grasp pose. And it makes fewer assumptions: thousands of part geometries (simulated with their millimeter clearances intact by our open-source generator, Section 2.2), randomized grasps, and starts centimeters from the hole. Prior systems have one or the other, and most have both (Section 4). Among [1–5], [1] adds tactile sensing, [2–4] estimate force from motor current or joint torque, and [5] gets the hole pose from a camera. [2–5] all insert round pegs of known size, rigidly mounted to the robot, and [2–4] each test only one peg.

This allows us to demonstrate three things that were not known to be possible, all with a single policy across thousands of part geometries. (1) Proprioception carries enough information to find the hole. With vision, a hole-pose prior, or a narrow task setup, proprioception only handles a final alignment or a fixed search. Here it must find the opening through contact across widely varied parts, grasps, and starts, and its actions respond to the contact state it infers (Section 3.1). (2) A behavior exists that needs no knowledge of the peg's pose in the hand, and our policy finds it. It tolerates grasps anywhere on the finger pads, tilted up to 20°, and in-hand slip: on the real robot, the peg slips a median of 4.2 mm from its initial grasp, several times the clearance (App. C). Slip is a leading real-world failure of insertion policies (Tang et al., 2023), and [2–5] avoid it only by mounting the peg rigidly. (3) Contact-rich search, inference, and insertion transfer zero-shot from simulation. Despite the well-known contact sim2real gap, the frozen policy inserts five real pegs of different sizes and shapes, including an unseen cross-section, in 74% of 151 trials, and a probe trained only in simulation reads the same contact modes from real joint histories (App. C). We have added [2–4] to Section 4.

  1. “Limited baselines and insufficient in-depth analysis. (...) Could the authors provide comparisons with stronger visual or force-based insertion policies, ideally including real-world evaluations?”

To address the concern, we ran new comparisons in the real world, on the same robot and with the same success criterion as Section 3.3. They represent two routes a practitioner today would take toward a stronger insertion policy: (1) training a BC policy with real-world collected demonstrations, (2) extending the sim-to-real recipe with additional modalities like camera images or F/T sensor readings. Both are new in the revision, with full implementation details in Appendices D and E; we summarize them here.

A. <u>BC on real-world demonstrations.</u> | We collected 59 teleoperated demonstrations and trained two policies on them: one on these demonstrations alone, and one fine-tuned from a VLA (π0.5). We made their task easier than the blind policy's on purpose: they are tested on the very peg they were shown (the 1.25 mm-clearance stadium of Fig. 10), always held in the same grasp, and they learn from real images, so no sim-to-real gap stands in their way (Appendix E).

PolicyTrained onTested onReal-world success
Behavior cloning (wrist camera)59 real world demos, single pegsame peg, fixed grasp10/33 (30%)
π0.5, full fine-tuning (wrist camera + robot state)59 real world demos, single pegsame peg, fixed grasp16/35 (46%)
Blind policy, ours (proprioception)RL in sim, random pegsame peg, randomized grasp24/30 (80%)

Even with in-domain data and a single geometry, both fall well short of the blind policy. We believe this reflects what demonstrations can teach: an operator who can see the hole shows how to reach it, not how to find it by touch.

B. <u>Policy with F/T and Vision alongside proprioception</u> | We kept training fixed and changed only what the actor observes, adding to proprioception either wrist force/torque or a wrist camera, encoded by a CNN trained from scratch or by frozen pretrained DINO features. Each was deployed zero-shot, as in Section 3.3 (Appendix D).

Actor observations (all trained in simulation, same recipe)Real-world successSim success on matched resets
Proprioception (ours)24/30 (80%)97%
Proprioception + wrist force/torque16/30 (53.3%)94.4%
Proprioception + wrist camera, CNN trained from scratch5/30 (16.7%)83.4%
Proprioception + wrist camera, frozen DINO features17/30 (56.7%)97.8%

Adding either sensor lowers real-world success: each new modality brings its own sim-to-real gap, rendered versus real images or simulated versus measured forces. Sim-to-real transfer is itself a distribution shift, and like the unseen geometries of Section 3.2, it rewards relying on signals that change little when the world does.

Together, these comparisons show that the blind policy's advantage is not an artifact of simple, simulated baselines.

  1. “While the paper demonstrates promising results, it provides limited mechanistic insight into why proprioception alone is sufficient for precise assembly and what drives its superior generalization.”
  • Why proprioception alone is sufficient for precise assembly? The discrepancy between the commanded and achieved joint position acts as an implicit contact sensor. The proprioception policy learns a strategy that relies on inferring and acting on the contact mode. Contact supervision explicitly incentivizes recovering the contact mode, which makes this behavior more easily accessible to the policy.
  • What drives proprioception’s superior generalization? The history latent is trained with both contact supervision and task reward for all policies. This means that any policy is free to learn a representation that is less reliant on the contact state if it leads to better task reward. The visual policy can achieve higher throughput and therefore higher reward by visually aligning the peg and hole, rather than searching for the hole through contact. However, these visual features are more vulnerable to held-out geometric variation (as one concrete example, imagine the peg is much smaller than expected, and an edge that was visible during training may become occluded by the gripper). The contact state representation that the proprioceptive policy learns to rely on, on the other hand, remains similar regardless of the object size, aspect ratio, etc.
Reviewer_daJmOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper learns peg insertion from joint-derived state and action history, without vision or force/torque measurements as policy inputs. Training uses a privileged critic and auxiliary contact labels in simulation. It also introduces mjfit, a procedural assembly-geometry generator for MuJoCo. A frozen policy is evaluated on a real arm across five pegs, with 112 successes in 151 reported attempts.

Critical strengths
  1. The real-robot result is interesting. A single policy completes insertion across several part sizes and an unseen cross-section using only joint-derived observations and its previous actions. Performance varies noticeably across parts, but the result is broader than a demonstration on one peg with a fixed grasp.

  2. mjfit is a useful contribution independently of the policy's comparison with vision.

  3. The paper makes an effort to understand the learned behavior. Contact-mode probing is better from the history latent than from the current observation, and swapping latents across modes changes actions more than swapping within a mode. Together, these results suggest that it is reacting to contact-related information rather than executing a fixed search trajectory.

Critical weaknesses

I feel the comparison does not support the broad claim that proprioception generalizes better than vision. The blind policy receives a privileged critic and auxiliary contact supervision during training. Section 2.3 says that the visual policy's critic sees only images, and no matching contact-supervision loss is described for that policy. Thus, the comparison changes training information as well as the actor's sensing modality.

The proposed history and contact-supervision design is not directly ablated. Probing shows that information can be decoded, and latent swapping shows that the policy uses its latent. Neither tells us whether the auxiliary contact objective is needed to learn a successful insertion policy. It would be helpful to show what happens when the contact loss is removed and the history encoder replaced with current observations.

Questions for authors

Does the unseen-shape gap remain when the visual policy has the same privileged critic and contact supervision? A vision-plus-proprioception comparison would also help distinguish the benefit of proprioceptive feedback from a supposed disadvantage of vision.

How do success and transfer change without the contact loss and without history? These direct ablations would be more convincing than further latent probes.

2 comments

We thank the reviewer for giving such constructive feedback. We address the concerns below:


  1. “The blind policy receives a privileged critic and auxiliary contact supervision during training.”

We agree that the “image-only” critic is a potential confounder, and have re-trained the camera-based policy, giving the critic access to the exact same privileged information as the proprioceptive policy (now, the two policies differ only in the actor’s observations) and updated Figs. 7 and 9, as well as Tab. A4, accordingly. The throughput (Fig. 7) and robustness (Fig. 9) trends both remains similar with the privileged critic. We have also clarified in the manuscript that the camera-based policy is indeed trained with the same contact-supervision loss as the proprioception policy.


  1. It would be helpful to show what happens when the contact loss is removed and the history encoder replaced with current observations."

To address this concern, we run two sets of ablation experiments (1) removing information from the observation with the same architecture, and (2) removing components from the architecture (contact supervision, temporal encoder) with the same observation.

  • In terms of removing information from the observation, we find that the biggest loss in performance comes from removing previous action and velocity observations, which deprive the policy of any information which can be used to infer the force. We also verify that the policy is not able to learn a purely open-loop solution (i.e. a “better spiral search”) when given only a clock as input. We added the experimental details in Appendix D1.
Actor inputBest success (%)
Proprio, 30-frame history (Ours)93.3
Proprio, 1 frame67.2
Joint + TCP position, 1 frame (no force information)7.2
Clock only (open loop)3.8
  • In terms of removing components from the architecture, both the contact supervision and temporal encoder play a large role in effectively compressing the history into an effective representation for the policy to use. We added the experimental details in Appendix D2.
PolicyBest success (%)
Encoder + supervision93.3
Encoder, no supervision42.0
No encoder (MLP), no supervision17.2

  1. "A vision-plus-proprioception comparison would also help distinguish the benefit of proprioceptive feedback from a supposed disadvantage of vision."

We have trained a vision-plus-proprioception policy and show that its performance falls between the proprio-only and vision-only policies on held-out geometries (Appendix D3, Figure 13). We hypothesize that this is because some portions of the task (i.e. finding the hole) are easier to solve with vision than proprio, so the fused policy learns to over-rely it. How to better combine multimodal inputs to promote generalization remains an exciting direction for future study.

Let us know if you have any further questions or clarifications.

Comments on the paper

1 comment
AuthorsOpenReview ↗

To All Reviewers

We thank all reviewers for their valuable feedback and constructive suggestions. We have edited the manuscript accordingly and marked all changes in blue. We provide individual responses to each reviewer below. Please let us know if any further clarification would be helpful in making your final decision.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.