FPO++ Code Release

[project page] [arxiv]

This repository contains experiments for FPO++ (Flow Policy Optimization), which have been cleaned up and refactored for release.

Flow Policy Gradients for Robot Control. Brent Yi12*, Hongsuk Choi12*, Himanshu Gaurav Singh12, Xiaoyu Huang12, Takara E. Truong13, Carmelo Sferrazza1, Yi Ma24, Rocky Duan1†, Pieter Abbeel12†, Guanya Shi15†, Karen Liu13†, and Angjoo Kanazawa12†

¹Amazon FAR, ²UC Berkeley, ³Stanford, ⁴HKU, ⁵CMU, ^*Equal Contribution, ^†Amazon FAR Team Co-lead

Directory Structure

• isaaclab_experiments/ — Isaac Lab experiments: velocity-conditioned locomotion (6+ robots). Includes training commands and expected training curves.
• manipulation_experiments/ — Manipulation experiments: pretraining and fine-tuning across five manipulation tasks using FPO++, Vanilla FPO, and DPPO variants. Includes training commands and expected training curves.

Each directory is self-contained with its own setup_env.sh, source_env.sh, and README with setup/usage instructions.

Setup

Each experiment directory has independent dependencies and conda environments. See the README in each directory for detailed setup instructions. In brief:

# Isaac Lab experiments
cd isaaclab_experiments
bash setup_env.sh        # one-time setup
source source_env.sh     # activate env (each session)

# Manipulation experiments
cd manipulation_experiments
bash setup_env.sh        # one-time setup
source source_env.sh     # activate env (each session)

Git Submodules

This repository uses git submodules for Isaac Lab extension (under isaaclab_experiments/). After cloning, initialize them with:

git submodule update --init --recursive

This is also done automatically by isaaclab_experiments/setup_env.sh.

Licenses

This repository includes and adapts code from the following third-party projects. Original license files and copyright headers are retained in all cases.

Isaac Lab Experiments (isaaclab_experiments/)

Project	License	Inclusion	What we use/adapt
IsaacLab	BSD-3-Clause	Git submodule	Simulation framework, task definitions, robot assets
rsl_rl	BSD-3-Clause	Adapted (not vendored)	Actor-critic modules, on-policy runner, rollout storage, normalizer

The isaaclab_fpo package adapts code from rsl_rl and IsaacLab (VecEnv wrapper, config dataclasses, training/evaluation scripts). Original copyright headers are retained in all adapted files.

Manipulation Experiments (manipulation_experiments/)

Project	License	Inclusion	What we use/adapt
LeRobot	Apache-2.0	Vendored under thirdparty/lerobot/ (minor modifications, see CHANGES.txt)	Policy architectures and training utilities
robosuite	MIT	Vendored under thirdparty/robosuite/ (minor modifications, see CHANGES.txt)	Simulation environments for manipulation tasks
DPPO	MIT	Adapted file (src/vit.py)	Vision Transformer encoder
DexMimicGen	NVIDIA Source Code License (non-commercial)	Cloned at setup time	Dexterous manipulation data generation