PaperlayerView exact revision on arXiv

Source-linked AI summary

AutoMate: Specialist and Generalist Assembly Policies over Diverse Geometries

Bingjie Tang, Iretiayo Akinola, Jie Xu, Bowen Wen, Ankur Handa, Karl Van Wyk, Dieter Fox, Gaurav S. Sukhatme, Fabio Ramos, Yashraj Narang

arXiv:2407.08028v2cs.RO

TL;DR

Robotic assembly in high-mixture settings requires adaptation to diverse parts and poses, while simulation-based development of generalist assembly policies had not been explored. AutoMate combines assembly data and simulation with specialist and generalist policy learning, reporting strong simulation results and zero-shot transfer to real-world assembly.

  • Problem

    Prior work largely trained specialist policies for 1–5 parts, with only some reaching 15–20, and simulation-based development and transfer of generalist assembly policies had not been explored.

    Source: §I · p. 1

  • Method

    AutoMate provides a 100-assembly dataset and parallelized simulators, combining assembly-by-disassembly, imitation-based reinforcement learning, and trajectory matching for specialists with geometric encoding, policy distillation, and curriculum-based RL fine-tuning for generalists.

    Source: §I · pp. 1–2

  • Results

    In simulation, specialists solved 80 assemblies at ≈80%+ success and a generalist solved 20 at 80%+; zero-shot success was 86.5% for specialists and 84.5% for the generalist.

    Source: §I · p. 2

  • Takeaways & Limitations

    AutoMate demonstrates a simulation-based framework for learning specialist and generalist policies across many assemblies and a system for zero-shot sim-to-real transfer over that range.

    Source: §I · p. 2

    Specialist policies achieved less than 80% success on 20 assemblies.

    Source: §VIII · p. 10

Abstract

from arXiv · show

Robotic assembly for high-mixture settings requires adaptivity to diverse parts and poses, which is an open challenge. Meanwhile, in other areas of robotics, large models and sim-to-real have led to tremendous progress. Inspired by such work, we present AutoMate, a learning framework and system that consists of 4 parts: 1) a dataset of 100 assemblies compatible with simulation and the real world, along with parallelized simulation environments for policy learning, 2) a novel simulation-based approach for learning specialist (i.e., part-specific) policies and generalist (i.e., unified) assembly policies, 3) demonstrations of specialist policies that individually solve 80 assemblies with 80% or higher success rates in simulation, as well as a generalist policy that jointly solves 20 assemblies with an 80%+ success rate, and 4) zero-shot sim-to-real transfer that achieves similar (or better) performance than simulation, including on perception-initialized assembly. The key methodological takeaway is that a union of diverse algorithms from manufacturing engineering, character animation, and time-series analysis provides a generic and robust solution for a diverse range of robotic assembly problems. To our knowledge, AutoMate provides the first simulation-based framework for learning specialist and generalist policies over a wide range of assemblies, as well as the first system demonstrating zero-shot sim-to-real transfer over such a range. For videos and additional details, please see our project website: https://bingjietang718.github.io/automate/

I. INTRODUCTION

High-mixture assembly requires adapting to diverse part geometries and poses, a challenge that remains open. AutoMate combines simulation-based policy learning with a real-world system for specialist and generalist assembly.

I. INTRODUCTION

  • AutoMate pairs a dataset of 100 assemblies with parallelized simulation environments and methods for learning specialist and generalist policies.The dataset and environments cover realistic, 3D-printable assembly problems.Source: §I · pp. 1–2
  • In simulation, specialists solved 80 assemblies at ≈80%+ success and a generalist solved 20 at 80%+; zero-shot success was 86.5% for specialists and 84.5% for the generalist.The zero-shot results cover 20 assemblies and 200 trials for each policy type.Source: §I · p. 2

II. RELATED WORKS

  • The related-work review covers assembly datasets, specialist-policy learning methods, generalist policy distillation, and sim-to-real transfer.It focuses on four building blocks and research areas for robotic assembly.Source: §II · p. 2

A. Assembly Datasets and Environments

  • AutoMate brings together a 100-assembly CAD dataset, parallelized simulation environments, and a corresponding real-world benchmarking kit.The authors describe these components as a unified dataset and environment for assembly sim-to-real transfer at appreciable scale.Source: A. Assembly Datasets and Environments · p. 2

B. Learning Specialist Assembly Policies

  • Specialist learning combines assembly-by-disassembly, RL with a path-imitation objective, and trajectory matching via dynamic time warping.The approach selects demonstrations by matching paths rather than states or state-action pairs.Source: B. Learning Specialist Assembly Policies · pp. 2–3

C. Learning Generalist Assembly Policies

  • Generalist learning distills multiple specialist policies through behavior cloning and DAgger, followed by RL fine-tuning, and deploys the resulting policy in the real world.The paper contrasts its assembly application and real-world deployment with prior multi-task distillation work focused on simulation.Source: C. Learning Generalist Assembly Policies · p. 3

D. Sim-to-Real Transfer for Assembly

AutoMate pairs a printable, interpenetration-free 100-assembly dataset and parallelized simulation with specialist and generalist learning methods, then applies sim-to-real methods without human demonstrations or policy adaptation. The task spans small, diverse parts under appreciable 6-DOF pose randomization, without part-specific fixtures or force-torque sensing.

D. Sim-to-Real Transfer for Assembly

  • AutoMate applies sim-to-real methods without human demonstrations or policy adaptation, while reducing setup effort through haphazard part placement, a second gripper for sockets, optimized grasps, and 6D pose estimation.The system also uses identical control gains and action scales across assemblies.Source: D. Sim-to-Real Transfer for Assembly · p. 3

III. PROBLEM DESCRIPTION

  • The target setting uses small parts with diverse geometries and appreciable 6-DOF pose randomization, without part-specific fixtures or a force-torque sensor.The stated assumptions include two-part assemblies, available part meshes, feasible grasps without regrasping, and approximately top-down insertion.Source: §III · p. 3

A. Assembly Dataset

  • 100 assemblies have interpenetration-free meshes and 1 mm diametral clearance for simulation, and can also be 3D-printed and assembled in reality.The real-world versions have 0.5–1.0 mm diametral clearance.Source: A. Assembly Dataset · p. 4

B. Assembly Environments

  • Parallelized environments cover all 100 assemblies, with a Franka robot, randomized initial robot and socket states, and a plug initially held by the robot.Initial states can be modified for custom applications.Source: B. Assembly Environments · p. 4

V. LEARNING METHODS

  • Specialist learning combines assembly-by-disassembly, imitation-guided reinforcement learning, and trajectory matching, while generalist learning combines geometric encoding, policy distillation, and curriculum-based fine-tuning.The paper notes that prior sim-to-real assembly studies often used large parts or clearances, specialized fixtures, human demonstrations, or real-world policy adaptation.Source: D. Sim-to-Real Transfer for Assembly · p. 3, §V · pp. 4–5

A. Specialist Learning: Assembly-by-Disassembly

Specialist training generates assembly demonstrations by reversing successful simulated disassembly paths, then augments reinforcement learning with an imitation reward. The approach uses grasp screening and trajectory demonstrations to support learning across randomized assembly starts.

A. Specialist Learning: Assembly-by-Disassembly

  • Assembly demonstrations are generated by reversing simulated disassembly paths, avoiding the challenges of collecting human demonstrations or planning narrow-passage assembly motions.Only paths, not velocities or accelerations, can be physically reversed.Source: A. Specialist Learning:… · p. 5
  • For each assembly, grasp candidates are screened with randomized-pose physics trials, and the highest-performing grasp is used to collect 100 successful disassembly paths.The grasp evaluation tests candidates over 1,000 trials before selecting the best-performing grasp.Source: A. Specialist Learning:… · p. 5

B. Specialist Learning: RL with Imitation Objective

  • The specialist policy is trained with PPO on randomized assembly states, using observations of joint angles and end-effector poses plus noisy socket-pose observations.Actions are incremental pose targets for a task-space impedance controller.Source: B. Specialist Learning: RL with… · p. 5
  • An imitation reward is added to the baseline reward, using the maximum per-timestep imitation reward across demonstrations for the assembly.The baseline and imitation terms are weighted by separate hyperparameters.Source: B. Specialist Learning: RL with… · p. 5

C. Specialist Learning: Trajectory Matching via Dynamic Time Warping and Signature Transforms

Specialist imitation compares the robot’s traversed path with reversed disassembly demonstrations using trajectory metrics designed to handle differing speeds and discretizations. Generalist training encodes part geometry, distills specialist policies, and fine-tunes with a pose curriculum.

C. Specialist Learning: Trajectory Matching via Dynamic Time Warping and Signature Transforms

  • Trajectory matching addresses the mismatch in speed and spatial sampling between the robot’s path and reversed disassembly demonstrations.The intended distance metric should be insensitive to speed or sampling rate.Source: C. Specialist Learning: Trajectory… · p. 6
  • Dynamic time warping aligns time-series points monotonically while minimizing summed Euclidean matching cost, and is applied to 10-step end-effector windows against demonstration segments.The compared demonstration segment is bounded by points closest to the window’s first and current end-effector positions.Source: C. Specialist Learning: Trajectory… · p. 6
  • Signature transforms represent 3D paths through path integrals and provide translation and reparameterization invariance that mitigate discretization sensitivity.The method compares signatures of the robot’s traversed path and a demonstration segment.Source: C. Specialist Learning: Trajectory… · p. 6

D. Generalist Learning: Geometric Encoding, Policy Distillation, and Curriculum-based RL Fine-tuning

  • A PointNet autoencoder maps plug and socket surface point clouds to latent geometry vectors, which are concatenated as generalist-policy inputs.The encoder is trained to minimize point-cloud reconstruction loss.Source: D. Generalist Learning: Geometric… · pp. 6–7
  • Generalist distillation uses behavior cloning followed by DAgger, then RL fine-tuning with a sampling-based curriculum that raises the lower bound on initial plug height across stages.The curriculum exposes the agent to the full initial-pose range at its start.Source: D. Generalist Learning: Geometric… · p. 7

VI. SPECIALIST AND GENERALIST POLICIES

AutoMate’s specialist and generalist evaluations test a unified approach to diverse assembly geometries. Specialist policies perform best with trajectory matching and consistently solve many assemblies, though the most challenging geometries remain difficult.

VI. SPECIALIST AND GENERALIST POLICIES

  • AutoMate combines assembly-by-disassembly, RL with an imitation objective, and trajectory matching via dynamic time warping and signature transforms.Evaluations use the maximum training bounds for initial-pose randomization and observation noise.Source: §VI · p. 7, A. Evaluations of Specialist Policies · p. 7

A. Evaluations of Specialist Policies

  • 81.50 ± 24.42% success for DTW and 77.46 ± 22.61% for signature transforms exceed 59.11 ± 19.65% for state-based matching and 38.45 ± 32.16% for IndustReal.DTW performs slightly better than signature transforms across all 100 assemblies.Source: A. Evaluations of Specialist Policies · p. 8
  • 81.50 ± 24.42% success for AutoMate exceeds 38.45±32.16% for IndustReal across all 100 assemblies.Go to Goal, Top Down, and Follow Trajectory score 15.44 ± 11.56%, 21.54 ± 15.49%, and 21.62±22.32%, respectively.Source: A. Evaluations of Specialist Policies · p. 8
  • For 80 assemblies, AutoMate achieves ≈80% (78%) success or higher, and for 55 assemblies it achieves 90% or higher.The bottom 20% most challenging assemblies tend to have small-diameter plugs, small socket contact surfaces, or stair-like internal socket features.Source: A. Evaluations of Specialist Policies · p. 8

B. Evaluation of Generalist Policy

AutoMate trains a generalist policy by distilling specialist policies, then refining the policy with DAgger and reinforcement learning. With sampling-based curriculum fine-tuning, the generalist reaches about 80% success across 20 assemblies; the paper also reports zero-shot real-world transfer outcomes comparable to or better than simulation.

B. Evaluation of Generalist Policy

  • SBC increases the lower bound of initial-pose randomization at each stage while keeping the upper bound fixed.The generalist training sequence uses behavior cloning, DAgger, RL fine-tuning, and sampling-based curriculum fine-tuning.Source: B. Evaluation of Generalist Policy · p. 8
  • 80.42 ± 20.93% success across 20 assemblies for BC + DAgger + RL with SBC exceeds 52.85 ± 14.01% without SBC and 48.43 ± 15.28% for RL with SBC from scratch.BC and BC + DAgger score 28.84 ±15.23% and 31.06 ± 15.06%, respectively.Source: B. Evaluation of Generalist Policy · p. 8, Figure 8 · p. 9

VII. SIM-TO-REAL TRANSFER

  • The authors report that specialist and generalist policies transfer zero-shot to real-world outcomes equivalent to, and sometimes better than, simulation.They attribute this supported conclusion to combining their learning approaches with sim-to-real transfer and pose-estimation methods.Source: §VII · pp. 8–9

A. Real-World System Design

The real-world system combines perception, grasp optimization, and learned assembly policies, with specialist and generalist performance close to simulation. The paper also reports perception-initialized results and identifies boundaries for the current assembly and trajectory-matching scope.

A. Real-World System Design

B. Real-World Policy Evaluations

C. Real-World Perception-Initialized Evaluation

VIII. CONCLUSIONS

  • Twenty assemblies have specialist success rates below 80%, with failures typically caused by unstable irregular-geometry grasps or sudden plug–socket slip.The authors anticipate higher-force grippers or recovery demonstrations may address these failures.Source: §VIII · p. 10
  • The current work solves two-part assemblies without sequence planning; an accelerated planner for multi-part assemblies is future work.Source: §VIII · p. 10
  • Trajectory matching currently focuses on R3 paths; the authors propose extending it to SE(3) for assemblies requiring significant reorientation.Source: §VIII · p. 10

B. Real-World Policy Evaluations

VIII. CONCLUSIONS

  • The generalist is distilled from 20 specialists, and the authors plan to distill from more specialists using larger-capacity architectures.Source: §VIII · p. 11

APPENDIX

The appendix sets out the real-world experimental setup and formalizes assembly as controlling a robot to move a plug from an initial pose to a constrained assembled pose.

A. Problem Description: Real-World Experimental Setup

B. Problem Description: Formal Problem Statement

C. Methods: Mesh Preprocessing

The methods prepare realistic meshes, select robust grasps through simulation, generate disassembly trajectories, and train policies with a reward-based reinforcement-learning formulation.

C. Methods: Mesh Preprocessing

  • Meshes are scaled, reoriented, translated, depenetrated to provide clearance, optionally chamfered, and subdivided for simulation contact resolution.The desired radial clearance is 0.5 mm; chamfered and unchamfered meshes are provided for simulation, while chamfered versions are 3D-printed for real-world use.Source: C. Methods: Mesh Preprocessing · p. 16

E. Methods: Disassembly Path Generation

  • Disassembly paths are generated in physics simulation and reversed for use in assembly.The paper visualizes 100 generated paths for an assembly with a deep socket.Source: Figure S13 · p. 17

D. Methods: Grasp Optimization

  • For each assembly, 100 candidate grasps are sampled and tested in 1,000 simulated lift-and-move trials; the highest-success grasp is selected.The resulting grasp is used for disassembly trajectory generation, policy training, and real-world deployment.Source: D. Methods: Grasp Optimization · pp. 16–17, Figure S12 · p. 17

F. Methods: Reinforcement Learning

  • The RL objective maximizes expected discounted reward over a horizon, combining task-difficulty and simulation-error weighting with distance, imitation, and success rewards.The policies are trained with PPO; the return weights task difficulty through SBC and simulation error through SAPU.Source: F. Methods: Reinforcement Learning · pp. 17–19

G. Methods: Dynamic Time Warping

Dynamic time warping selects a disassembly path resembling the robot's traversed path by matching points in two sequences under endpoint and monotonicity constraints.

G. Methods: Dynamic Time Warping

  • At each timestep, DTW selects the closest disassembly path to the assembly path already traversed by the robot.Source: G. Methods: Dynamic Time Warping · p. 19
  • DTW minimizes summed point-to-point distance while aligning the first and last sequence points and preserving monotonic matches.Each point may match one or more points in the other sequence.Source: G. Methods: Dynamic Time Warping · p. 19
  • The dynamic-programming matrix accumulates the current point distance plus the minimum cost from three predecessor matches.The final matrix entry gives the minimum accumulated matching cost between the sequences.Source: G. Methods: Dynamic Time Warping · pp. 19–20

H. Methods: Signature Transform

AutoMate uses path signatures to compare the robot’s traversed assembly path with reversed disassembly paths and select the closest one to imitate. The representation summarizes path integrals and is invariant to translation and reparameterization.

H. Methods: Signature Transform

  • At each timestep, the method selects the reversed disassembly path whose signature is closest to the robot’s traversed path.It compares path signatures using their L2 distance.Source: H. Methods: Signature Transform · p. 20
  • A path signature is an ordered collection of path integrals, with successive levels encoding interactions among more path coordinates.For a 3D path, the first level has 3 integrals and the second has 9.Source: H. Methods: Signature Transform · p. 20
  • Translation and reparameterization invariance make signatures suitable for paths with offsets or different speeds and sampling schemes.The paper uses the discrete-time form because its data are discrete.Source: H. Methods: Signature Transform · p. 20

I. Methods: Point-Cloud Autoencoder

The final specialist approach outperforms Follow Trajectory on nearly all assemblies. The section also describes a geometry autoencoder and tests specialist robustness to pose randomization and observation noise, including fine-tuning under stronger shifts.

I. Methods: Point-Cloud Autoencoder

  • The generalist’s geometry input is a latent vector learned by encoding point clouds and reconstructing them with a decoder.The autoencoder minimizes Chamfer distance; the paper suggests explicit or implicit surface information as future work that may improve generalist success.Source: §I · p. 21, Figure S14 · p. 21

J. Results: Specialist Policies (Continued)

K. Results: Robustness of Specialist Policies to Initial-Pose Randomization in Simulation

  • Specialists maintain high performance in-distribution and moderately degrade under larger-than-training pose randomization; fine-tuning with twice the original randomization is effective.The evaluation covers 10 assemblies, with quantitative results reported in Table VI.Source: K. Results: Robustness of Specialist… · p. 22, Table VI · p. 24

L. Results: Robustness of Specialist Policies to Observation Noise in Simulation

  • Specialists outperform Follow Trajectory across assemblies under socket-pose observation noise, with moderate degradation out-of-distribution, especially near critical part length scales.Fine-tuning with twice the original observation noise can moderately improve performance under higher noise.Source: L. Results: Robustness of Specialist… · pp. 22, 24

M. Results: Generalist Policies

Generalist success is high when trained from 10 or 20 specialists but drops as the specialist set grows. The section also describes camera calibration and a CAD-based pose-estimation pipeline for deploying policies trained from simulated 6D poses.

M. Results: Generalist Policies

  • Generalist success is approximately 80% with 10 or 20 specialists, then drops to approximately 55% with 30, 30% with 40, and 20% with more specialists.Batches are evenly sampled in t-SNE space, and each generalist is evaluated on its corresponding assemblies over 5,000 trials.Source: M. Results: Generalist Policies · p. 24, Figure S22 · p. 25

1) Camera Calibration and Tuning:

  • Diligent calibration and tuning of a single wrist-mounted RGB-D camera improve image and point-cloud quality enough for high-precision assembly of parts smaller than the robot manipulator.The camera is an Intel RealSense D435.Source: 1) Camera Calibration and Tuning · p. 24

2) Pose Estimation:

  • The real-world pipeline uses RGB-D images, user-selected part pixels, segmentation, and a known CAD model to estimate each part’s 6D pose for the policy.The paper assumes a known CAD model for each part; the described pipeline takes approximately 20 seconds.Source: 2) Pose Estimation · pp. 24–26

O. Results: Robustness of Sim-to-Real Transfer to Observation Noise

Real-world assembly remains successful under the tested image and control perturbations. Specialist policies moderately degrade when initial-pose randomization exceeds the training range.

O. Results: Robustness of Sim-to-Real Transfer to Observation Noise

  • The perception-initialized task achieves high success despite pose-estimator inaccuracies, and the estimator remains robust to brightness, contrast, and Gaussian image noise.The qualitative estimator test covers five assemblies.Source: O. Results: Robustness of Sim-to-Real… · p. 26
  • With added control noise of ±2 mm per axis and ±5° per rotation at every timestep, the system maintains high success across five assemblies.The paper anticipates degradation at higher noise levels but says those levels may be unrealistic for its hardware and application.Source: O. Results: Robustness of Sim-to-Real… · p. 26, Table VIII · p. 26

P. Results: Robustness of Sim-to-Real Transfer to Initial-Pose Randomization

Loading 2407.08028v2…