SERL: A Software Suite for Sample-Efficient Robotic Reinforcement Learning

ICRA 2024

Featured on Hugging Face Daily Paper

^*Equal Contribution; ¹Department of EECS, University of California, Berkeley; ²Department of Computer Science, University of Washington; ³Department of Computer Science, Stanford University; ⁴Intrinsic Innovation LLC

SERL is a ready-to-use software suite for robotic RL, featuring sample efficient off-policy algorithms, various reward specification methods, and advanced controller for popular robots. It includes example tasks such as PCB assembly, cable routing, and reset-freeobject relocation. Remarkably, it trains policies in just 25 to 50 minutes, outperforming previous benchmarks with high success rates and robustness.

PCB Component Insertion

The agent successfully inserts the PCB component as trained.

The agent successfully inserts the PCB component after being blind-folded and the board moved.

The agent successfully inserts the PCB component despite multiple distractor objects on the PCB board.

The agent succeeds when the board is not fixed to the table.

The agent successfully inserts the PCB component into a the board in a different pose than training.

The agent successfully finds the holes in the board after it is moved.

The agent successfully finds the holes in the board after it is continuously moved.

Cable Routing

The agent successfully routes the cable as trained.

The agent generalizes to route the cable through an unseen clip pose.

The agent successfully routes the cable despite continuous perturbation to the clip.

The agent generalizes to route the cable through an unseen clip pose.

Object Relocation

The agent relocates the object as trained.

The agent relocates the trained object despite multiple distractors in the bins.

The agent relocates the trained object despite a distractor in the scene.

The agent generalizes to relocate a different object than trained.

The agent completes the task as trained.

BibTeX

@misc{luo2024serl, title={SERL: A Software Suite for Sample-Efficient Robotic Reinforcement Learning}, author={Jianlan Luo and Zheyuan Hu and Charles Xu and You Liang Tan and Jacob Berg and Archit Sharma and Stefan Schaal and Chelsea Finn and Abhishek Gupta and Sergey Levine}, year={2024}, eprint={2401.16013}, archivePrefix={arXiv}, primaryClass={cs.RO} }