boost-bc-flow-policies
Flow-matching behavior policies for robotics
2 min readupdated
This repository provides a collection of fixed baseline models (behavior-cloning policies) with varying success rates (from 10% to 90%), trained via flow-matching across different robotics simulations. The models cover cube manipulation tasks in MimicGen, tool hanging in robosuite, furniture assembly in FurnitureBench, and pen manipulation in Adroit Pen. They are designed as stable starting points for subsequent reinforcement learning fine-tuning (RL fine-tuning). These models will be useful for researchers and engineers working on robotics and manipulator simulation.
Key features
- Ready-to-use checkpoints with success rate gradients (10, 25, 50, 75, 90%) for five complex robotics tasks.
- Flow-matching architecture utilizing ResNet-18 encoders for processing camera images and 1D temporal U-Net networks.
- Precisely fixed demonstration sets for each task, which is critical for a correct start for RL algorithms.
- A complete set of configuration files, including EMA actor weights (ema_actor.pt), normalization parameters, and deployment metadata.
What you can do
- Use as a base for RL: apply checkpoints with different success rates to jumpstart reinforcement learning on the same demonstration datasets.
- Run evaluation: test policy performance in MimicGen, robosuite, FurnitureBench, or Adroit Pen environments using the provided configuration files.
- Explore success rate impact: test how the quality of the base policy (e.g., 25% vs. 75%) affects final RL fine-tuning speed and stability.
Example
The `rfm_rl` codebase is used to launch and test policies. Each task folder contains an `ema_actor.pt` weight file, a `config.yaml` configuration file, and `deploy_metadata.json` metadata describing observation and action keys.
Read more
DetailsBenchmarks: task execution success rates from 10% to 89-93% depending on the checkpoint level and environment (MimicGen Stack, robosuite Tool Hang, FurnitureBench, Adroit Pen).
Architecture: image-based tasks use ResNet-18 + spatial softmax with a U-Net flow head (about 26.26 million parameters), while Pen uses a residual MLP flow head (about 8.71 million parameters).
Training: AdamW optimizer with adaptive EMA 0.9999. Weights inherit the licenses of the original datasets (MimicGen, robomimic, FurnitureBench, D4RL).










Comments