Lei-Kun/RL-100

What it solves

RL-100 provides a unified framework for improving robotic manipulation policies. It addresses the gap between initial imitation learning (behavior cloning) and high-performance, reliable real-world deployment by providing a repeatable pipeline for post-training policies using reinforcement learning (RL).

How it works

The system implements a multi-stage pipeline that moves from static data to active interaction:

1. Behavior Cloning: Initializes a policy from human demonstrations.
2. Offline RL: Refines the policy using logged trajectories via policy-gradient updates.
3. Online RL: Further fine-tunes the policy through real-world or simulated interactions.
4. One-Step Distillation: Converts multi-step diffusion or flow policies into efficient one-step policies (e.g., Diffusion-to-CM) for fast real-robot inference.
5. Data Flywheel: An iterative loop where offline-trained policies are deployed to collect new robot rollouts, which are then merged back into the training dataset to further improve the next round of offline RL.

Who it’s for

It is designed for robotics researchers and engineers working on visuomotor policies who need a standardized way to move from teleoperation data to robust, deployment-ready robotic control.

Highlights

• Flexible Policy Backbones: Supports both Diffusion and Flow policies.
• Multimodal Observations: Compatible with 3D point clouds and 2D RGB images.
• Versatile Control: Supports both action chunking and high-frequency single-action control.
• Integrated Data Tools: Includes utilities for teleoperation data collection and iterative Zarr dataset management.