HIL-SERL

Berkeley's real-robot RL system where a human can take over anytime, learning fine manipulation in 1 to 2.5 hours.

HIL-SERL is a real-robot visual reinforcement-learning system from Jianlan Luo, Charles Xu, Jeffrey Wu, and Sergey Levine at UC Berkeley, posted to arXiv in October 2024 and featured on the cover of Science Robotics in August 2025. It builds on the same group's open-source SERL software stack. First, teleoperation collects a handful of successful and failed examples to train a binary classifier as a sparse reward signal; a small number of demonstrations are then loaded into a replay buffer; online training uses an off-policy algorithm based on RLPD, which can reuse old data repeatedly, and an operator can take over and correct the robot at any time with a device like a SpaceMouse, with that correction data also fed back into learning. Real-robot RL is usually too slow and unstable to be practical; HIL-SERL uses a pretrained vision backbone, demonstrations plus human corrections, and a safe low-level controller to cut training down to 1 to 2.5 hours, reaching near-100% success on most tasks — roughly double the success rate and 1.8x the execution speed of imitation-learning baselines.

ExampleTasks in the paper include Jenga extraction — the arm whips a block out of a tower — flipping objects in a pan, and precision assembly tasks like a circuit board, an IKEA shelf, a car dashboard, and a timing belt.

Also called

Human-in-the-Loop Sample-Efficient Robotic Reinforcement Learning, Precise and Dexterous Robotic Manipulation via Human-in-the-Loop RL

Related

SERL · Human-in-the-Loop · Real-World Reinforcement Learning · Human-Gated DAgger · Reward Model · Reinforcement Fine-Tuning (RL Fine-Tuning)

Sources

Precise and Dexterous Robotic Manipulation via Human-in-the-Loop Reinforcement Learning (arXiv 2410.21845)
HIL-SERL 项目主页 (Chinese)
Jianlan Luo 个人主页（HIL-SERL 发表于 Science Robotics 2025） (Chinese)

As of

2025-08