FreeBAR

CoRL 2026 | Paper

This repository contains the official implementation of FreeBAR, an inference-time method for delay-aware execution of chunked flow-matching policies in real-time robotic control.

Overview

Real-time action chunking policies commit to a chunk before the next one is ready, so a new chunk is always planned against actions that are already being executed. FreeBAR resolves this at inference time by splitting the planning horizon into three regions and treating each one differently:

   committed prefix        near future              far future
   [0, d)                  [d, h)                   [h, H)
   ─────────────────       ─────────────────        ─────────────────
   strict alignment        soft alignment           unconstrained
   to committed actions    with exponential decay   (left free)

At every denoising step FreeBAR adds a control-barrier-function style guidance term to the flow velocity: the committed prefix is aligned strictly to the delayed actions, the near future is aligned softly so the chunk stays coherent across the boundary, and the far future is left free so the policy can still replan. The guidance multiplier λ comes from a closed-form QP solution, clipped to [0, max_guidance_weight].

FreeBAR is evaluated in two settings, both included here:

Setting	Policy backbone	Environment	Entry point
Kinetix	Flow policy (JAX)	Kinetix	kinetix/src/eval_flow_rtc.py
RoboCasa	GR00T N1.5	RoboCasa	Isaac-GR00T/scripts/run_eval_async.py

Repository Structure

FreeBar/
├── kinetix/                    # Kinetix implementation (JAX)
│   ├── src/
│   │   ├── model.py            # FlowPolicy + freebar_action
│   │   ├── eval_flow_rtc.py    # Evaluation entry point (naive / rtc / freebar)
│   │   └── train_flow.py       # Behavior-cloning training
│   ├── worlds/l/               # 12 Kinetix task levels
│   ├── third_party/kinetix/    # Vendored Kinetix environment
│   └── pretrained/             # Place checkpoints here (see pretrained/README.md)
│
├── Isaac-GR00T/                # GR00T fork with the FreeBAR action head
│   ├── gr00t/model/action_head/
│   │   ├── freebar.py          # FreeBAR guidance implementation
│   │   ├── freebar_recorder.py # Optional λ / gradient recording
│   │   └── flow_matching_action_head.py
│   └── scripts/run_eval_async.py
│
└── robocasa/                   # RoboCasa fork used by the GR00T evaluation
    ├── robocasa/
    ├── robomimic/
    └── scripts/                # ACT baselines (train / eval / convert)

What is not included

To keep the repository small, the following are intentionally excluded:

Excluded	How to obtain
robocasa/models/	Restore from upstream RoboCasa — see RoboCasa
RoboCasa textures and objects (~10 GB)	Downloaded by download_kitchen_assets
Behavior-cloning checkpoints	Train with kinetix/src/train_flow.py, or place your own under kinetix/pretrained/bc/
GR00T policy checkpoints	Pass your checkpoint path via --model_path

Installation

Environment

Environment Note: The Kinetix evaluation was developed and tested on Linux with CUDA. kinetix/pyproject.toml pins jax[cuda12], which has no macOS wheels, so uv sync will fail on macOS and Windows.

Kinetix

cd kinetix
uv sync

Place the behavior-cloning policy checkpoints under kinetix/pretrained/bc. The directory should contain numbered checkpoint folders with policy files, for example pretrained/bc/24/policies/worlds_l_grasp_easy.pkl. See kinetix/pretrained/README.md for the expected layout.

RoboCasa

RoboCasa needs two things that are not vendored in this repository:

• the upstream robosuite package, and
• the robocasa/models/ directory, which is excluded here to keep the release small (see below).

Everything else under robocasa/ is RoboCasa at commit 9a3a786, with one local change: the upstream numpy version assertion in robocasa/__init__.py is relaxed.

Set both up before installing the local packages:

# 1. robosuite (external dependency, installed next to this repository)
cd ..
git clone https://github.com/ARISE-Initiative/robosuite.git
cd robosuite
pip install -e .

# 2. restore robocasa/models/ from upstream RoboCasa
cd ..
git clone https://github.com/robocasa/robocasa.git robocasa-upstream
cd robocasa-upstream
git checkout 9a3a786          # the revision this fork is based on
cp -R robocasa/models ../FreeBar/robocasa/robocasa/

# 3. install the packages in this repository
cd ../FreeBar
pip install -e robocasa
pip install -e robocasa/robomimic
pip install -e Isaac-GR00T

# 4. configure RoboCasa paths and download the benchmark assets
python -m robocasa.scripts.setup_macros
python -m robocasa.scripts.download_kitchen_assets

robocasa/models/ is required at import time, so step 2 is not optional: robocasa.environments imports robocasa.models.fixtures, robocasa.models.objects and robocasa.models.scenes, and download_kitchen_assets.py reads models/assets/box_links/box_links_assets.json as soon as it starts. Without that directory neither the environments nor the asset downloader will run.

The asset download itself is about 10 GB of textures, objects and fixture meshes. Keep the downloaded assets outside Git; they are not part of this release either.

Usage

Kinetix

Run the default baseline comparison over naive and rtc:

cd kinetix
uv run src/eval_flow_rtc.py --run-path ./pretrained/bc

Run FreeBAR explicitly:

uv run src/eval_flow_rtc.py \
  --run-path ./pretrained/bc \
  --methods freebar

--methods accepts any of naive, rtc, freebar. Other useful flags are --quick-test, --test-pairs, --output-dir, --config.step, --config.num-evals and --config.num-flow-steps.

RoboCasa

Run FreeBAR:

cd Isaac-GR00T
MUJOCO_GL=egl python scripts/run_eval_async.py \
  --model_path /path/to/checkpoint-60000 \
  --method freebar \
  --n_action_steps 6 \
  --inference_delay 3 \
  --override_action_horizon 16 \
  --task_set atomic_seen \
  --split target \
  --use_chunked_execution \
  --n_envs 5 \
  --cbf_alpha 20.0 \
  --cbf_max_guidance_weight 10.0

RoboCasa keeps the public method set to naive and freebar. In comparison mode the selected --method is always recorded, and the default extra baseline is naive:

MUJOCO_GL=egl python scripts/run_eval_async.py \
  --model_path /path/to/checkpoint-60000 \
  --method freebar \
  --compare_methods \
  --n_action_steps 6 \
  --inference_delay 3 \
  --override_action_horizon 16 \
  --task_set atomic_seen \
  --split target \
  --use_chunked_execution

Citation

If you find this work useful, please cite:

Coming soon

📄 License

FreeBAR additions and top-level packaging are released under the Apache-2.0 License — see the LICENSE file for details. Third-party components retain their original licenses; see THIRD_PARTY_NOTICES.md.

📧 Contact

For questions and feedback, please open an issue or contact .