IndustReal: Transferring Contact-Rich Assembly Tasks from Simulation to Reality

TL;DR

IndustReal tackles sim-to-real transfer for contact-rich assembly by integrating three sim-based RL innovations—simulation-aware policy updates, SDF-based dense rewards, and sampling-based curricula—with a deployment-time policy-level action integrator. The approach is validated on Pegs/Holes, Gears/Gearshafts, and Connectors/Receptacles tasks, achieving high success rates in both simulation (82-99%) and real-world trials (83-99%) across hundreds of trials, without real-world policy adaptation. A complete end-to-end system is demonstrated, including perception, grasping, alignment, and insertion, and the authors provide IndustRealKit/IndustRealLib to encourage reproducibility. This work sets a new benchmark for RL-based, end-to-end sim-to-real assembly and provides a practical pathway toward reusable, hardware-enabled assembly pipelines.

Abstract

Robotic assembly is a longstanding challenge, requiring contact-rich interaction and high precision and accuracy. Many applications also require adaptivity to diverse parts, poses, and environments, as well as low cycle times. In other areas of robotics, simulation is a powerful tool to develop algorithms, generate datasets, and train agents. However, simulation has had a more limited impact on assembly. We present IndustReal, a set of algorithms, systems, and tools that solve assembly tasks in simulation with reinforcement learning (RL) and successfully achieve policy transfer to the real world. Specifically, we propose 1) simulation-aware policy updates, 2) signed-distance-field rewards, and 3) sampling-based curricula for robotic RL agents. We use these algorithms to enable robots to solve contact-rich pick, place, and insertion tasks in simulation. We then propose 4) a policy-level action integrator to minimize error at policy deployment time. We build and demonstrate a real-world robotic assembly system that uses the trained policies and action integrator to achieve repeatable performance in the real world. Finally, we present hardware and software tools that allow other researchers to reproduce our system and results. For videos and additional details, please see http://sites.google.com/nvidia.com/industreal .

Figures (16)

• Figure 1: Overview. Top: Simulation-based policy learning for one of our tasks, gear assembly. Middle: Proposed algorithms to facilitate sim-based learning and real-world deployment. Bottom: Successful transfer to the real world.
• Figure 2: Problem setup and decomposition. Column 1: Three types of assemblies. Columns 2-4: Goal states of Pick, Place, and Insert phases.
• Figure 3: Evaluation of Simulation-Aware Policy Update. Success rates are computed for episodes where the maximum interpenetration distance was less than the specified value at test time. Boxes indicate median and IQR.
• Figure 4: Joint evaluation of Simulation-Based Policy Update, SDF-Based Dense Reward, and Sampling-Based Curriculum. (A) Pegs and Holes assembly Insert policy. (B) Gears and Gearshafts assembly Insert policy. Engage denotes partial insertion. Policies were trained with moderate randomization (plug/socket randomization of $\pm$ 10 mm and 10 cm, respectively, and observation noise of $\pm$ 1 mm); thus, plots evaluate in-distribution and OOD performance.
• Figure 5: Evaluation of PLAI in simulation. Results of Nominal are annotated when outside of plot bounds. Full-axis plot is in \ref{['fig:control-ablation-full-appendix']}.