amazon-far/fpo-control

What is this?

FPO++ Code Release is the official code accompanying the paper Flow Policy Gradients for Robot Control. It provides reproducible experiments for the Flow Policy Optimization (FPO++) algorithm, a reinforcement‑learning method for training robot control policies.

Main Components

Component	Purpose	Key Dependencies
`isaaclab_experiments/`	Velocity‑conditioned locomotion experiments on 6+ robot platforms. Includes training scripts, expected learning curves, and all the environment setup needed to run the tasks.	Isaac Lab (simulation framework), rsl_rl (actor‑critic and on‑policy runner)
`manipulation_experiments/`	Pre‑training and fine‑tuning on five dexterous manipulation tasks. Provides scripts for FPO++, vanilla FPO, and DPPO baselines.	LeRobot (policy architectures), robosuite (manipulation simulators), DPPO (Vision Transformer encoder), DexMimicGen (data generation)

How to get started

Clone the repo (including submodules):

git clone --recursive https://github.com/amazon-far/fpo-control.git
cd fpo-control

Choose an experiment set and run its setup script:

# Locomotion (Isaac Lab)
cd isaaclab_experiments
bash setup_env.sh   # creates a conda env and installs dependencies
source source_env.sh

# Manipulation
cd ../manipulation_experiments
bash setup_env.sh
source source_env.sh

Follow the README inside the chosen directory for the exact training command (e.g., python train.py --config ...). The README also lists the expected training curves so you can verify that your run matches the paper.

Licensing & Third‑Party Code

The repository bundles several external projects, each kept under its original license:

Isaac Lab (BSD‑3) – simulation core and robot assets.
rsl_rl (BSD‑3) – RL utilities adapted for FPO.
LeRobot (Apache‑2.0) – policy network definitions.
robosuite (MIT) – manipulation environments.
DPPO (MIT) – Vision Transformer encoder.
DexMimicGen (NVIDIA Source Code License, non‑commercial) – data generation for dexterous hands. All copyright notices are retained in the adapted files.

Who is behind it?

The work is a collaboration between Amazon FAR, UC Berkeley, Stanford, HKU, and CMU. Lead authors are Brent Yi and Hongsuk Choi (equal contribution), with senior supervision from Pieter Abbeel, Guanya Shi, Karen Liu, and Angjoo Kanazawa.

When would you use this?

Research: Re‑implement or extend the FPO++ algorithm for new robot tasks.
Benchmarking: Compare FPO++ against vanilla FPO or DPPO on locomotion or manipulation benchmarks.
Education: Study a complete RL pipeline (environment, policy, training loop) in a realistic robotics simulator.

Quick take‑away

FPO++ Code Release is a genuine, research‑grade software package that lets you run state‑of‑the‑art reinforcement‑learning experiments for robot locomotion and dexterous manipulation, built on top of well‑known simulation and RL libraries.