BrainCo Isaac Lab

Isaac Lab environments, robot assets, and pretrained checkpoints for BrainCo dexterous manipulation tasks. Including:

• Revo3 in-hand repose
• Revo3 in-hand reorientation
• Revo3 HORA/RMA in-hand ball and cylinder rotation
• Revo3 right-hand lift
• RevoTron dynamic handover

Included Tasks

Robot	Framework	Public task ID	Checkpoint	Demo
Revo3	Direct	BrainCo-Direct-Revo3-Repose-Cube-v0	checkpoints/BrainCo-Direct-Revo3-Repose-Cube-v0.pt
Revo3	Direct	BrainCo-Direct-Revo3-Reorient-Cylinder-v0	checkpoints/BrainCo-Direct-Revo3-Reorient-Cylinder-v0.pt
Revo3	Dexsuite	BrainCo-Dexsuite-Revo3-Right-Lift-v0	checkpoints/BrainCo-Dexsuite-Revo3-Right-Lift-v0.pt
Revo3	HORA	BrainCo-Direct-Revo3-HoraRotate-Ball-v0	checkpoints/hora/revo3_right_ball_stage1_best.pth
Revo3	HORA	BrainCo-Direct-Revo3-HoraRotate-Cylinder-v0	checkpoints/hora/revo3_right_cylinder_stage1_best.pth
RevoTron	Dynamic Handover	BrainCo-Dynamic-Handover-Revo3-Cube-v0	checkpoints/dynamic_handover/revotron_handover.pth

Repository Layout

BrainCo-IsaacLab/
├── assets/                 # Robot assets, object meshes, and grasp caches
│   ├── usd/                # USD assets grouped by task family
│   └── urdf/               # URDF assets and meshes
├── checkpoints/            # Pretrained checkpoints
├── image/                  # Task demo GIFs
├── scripts/rsl_rl          # RSL-RL training and evaluation scripts
├── scripts/rl_games        # RL-Games training and evaluation scripts
├── scripts/hora            # HORA PPO / ProprioAdapt training and export scripts
├── deploy/revo3            # Revo3 sim-to-real deployment package
└── source/BrainCo_DexHand/ # Isaac Lab extension package

Requirements

• Python 3.10+
• A working Isaac Lab installation

This repository is distributed as an Isaac Lab extension package, not as a standalone simulator fork.

Installation

1. Install Isaac Lab by following the official Isaac Lab setup for your target version.
2. Clone this repository into your workspace.
3. Install the BrainCo extension:

cd source/BrainCo_DexHand
pip install -e .

Downloading Checkpoints

We provide a script to easily download all the pretrained checkpoints from our OSS server. Run the following command from the repository root:

./scripts/download-checkpoints.sh

This will download the *.pt files directly into the repository's checkpoints/ directory.

In-hand Manipulation

The in-hand manipulation environments use RSL-RL. The environments register when BrainCo_DexHand is importable in your Isaac Lab Python environment.

Check registration:

python -c "import BrainCo_DexHand"

Train:

python  scripts/rsl_rl/train.py --task BrainCo-Direct-Revo3-Repose-Cube-v0 --num_envs 8192 --headless
python  scripts/rsl_rl/train.py --task BrainCo-Direct-Revo3-Reorient-Cylinder-v0 --num_envs 4096 --headless

Evaluate:

python  scripts/rsl_rl/play.py --task BrainCo-Direct-Revo3-Repose-Cube-v0 --checkpoint checkpoints/BrainCo-Direct-Revo3-Repose-Cube-v0.pt --num_envs 1
python  scripts/rsl_rl/play.py --task BrainCo-Direct-Revo3-Reorient-Cylinder-v0 --checkpoint checkpoints/BrainCo-Direct-Revo3-Reorient-Cylinder-v0.pt --num_envs 1

Grasping

The grasping task uses the Dexsuite Revo3 environment with RSL-RL.

Train:

python  scripts/rsl_rl/train.py --task BrainCo-Dexsuite-Revo3-Right-Lift-v0 --num_envs 4096 --headless

Evaluate:

python  scripts/rsl_rl/play.py --task BrainCo-Dexsuite-Revo3-Right-Lift-Play-v0 --checkpoint checkpoints/BrainCo-Dexsuite-Revo3-Right-Lift-v0.pt --num_envs 1

HORA Training and Export

The HORA rotation tasks are integrated as a separate training path because their PPO, ProprioAdapt Stage 2 adapter, and checkpoint format are not RSL-RL-compatible. This integration is adapted from HORA.

Stage 1 PPO teacher training:

python scripts/hora/train.py --algo PPO --task ball --num_envs 16384 --headless
python scripts/hora/train.py --algo PPO --task cylinder --num_envs 16384 --headless

Stage 2 ProprioAdapt training from a Stage 1 checkpoint:

python scripts/hora/train.py --algo ProprioAdapt --checkpoint outputs/hora/revo3_right/run_ball/stage1_nn/best.pth --task ball --num_envs 16384 --headless
python scripts/hora/train.py --algo ProprioAdapt --checkpoint outputs/hora/revo3_right/run_cylinder/stage1_nn/best.pth --task cylinder --num_envs 16384 --headless

Evaluate a checkpoint:

python scripts/hora/train.py --algo PPO --task ball --num_envs 32 --test --checkpoint checkpoints/hora/revo3_right_ball_stage1_best.pth
python scripts/hora/train.py --algo PPO --task cylinder --num_envs 32 --test --checkpoint checkpoints/hora/revo3_right_cylinder_stage1_best.pth

Export a Stage 2 checkpoint to ONNX:

python scripts/hora/export_onnx.py \
  --stage stage2 \
  --checkpoint outputs/hora/revo3_right/run_ball/stage2_nn/model_best.ckpt \
  --output ball_hora_stage2.onnx

Dynamic Handover

Dynamic handover uses the RevoTron asset and RL-Games. This task is adapted from dynamic_handover. Run from the repository root with the Isaac Lab Python environment activated.

Train:

python scripts/rl_games/train.py \
  --task BrainCo-Dynamic-Handover-Revo3-Cube-v0 \
  --num_envs 4096 \
  --headless

Evaluate:

python scripts/rl_games/play.py \
  --task BrainCo-Dynamic-Handover-Revo3-Cube-Play-v0 \
  --num_envs 1 \
  --checkpoint checkpoints/dynamic_handover/revotron_handover.pth

To manually switch toss direction during play, pass a command file:

python scripts/rl_games/play.py \
  --task BrainCo-Dynamic-Handover-Revo3-Cube-Play-v0 \
  --num_envs 1 \
  --checkpoint checkpoints/dynamic_handover/revotron_handover.pth \
  --command-file /tmp/handover_command.txt

Then update the command from another terminal:

echo 'right_throw' > /tmp/handover_command.txt
echo 'left_throw' > /tmp/handover_command.txt

Sim-to-real Deployment

Below is a sim-to-real deployment of BrainCo-Direct-Revo3-HoraRotate-Cylinder-v0. More tasks' sim-to-real deployments will be released in the future.

The Revo3 deployment runtime lives in deploy/revo3. It is a lightweight, ROS-free package for running exported ONNX policies on the real Revo3 hand through the Revo3 Python SDK.

The deploy loop is closed-loop: it reads measured joint positions from hardware, builds the policy observation history, runs ONNX inference, maps policy-order targets to SDK motor order, and sends MIT commands back to the hand. It is not a fixed trajectory replay tool.

Install the deploy package:

cd deploy/revo3
pip install -e .

Install the hardware extra when running on the real hand:

pip install -e ".[hardware]"

The runtime uses:

• policy.onnx: exported policy network
• policy.yaml: policy I/O contract and metadata
• config/revo3_right.yaml: joint order, joint limits, SDK settings, MIT gains, and sim-to-real offsets

Run a dry-run policy loop without sending commands:

python scripts/run_policy.py \
  --onnx artifacts/repose_cube/policy.onnx \
  --policy artifacts/repose_cube/policy.yaml \
  --profile config/revo3_right.yaml \
  --dry-run

See deploy/revo3/README.md for export and hardware run commands.

Notes

• The Revo3 tasks IDs follow the BrainCo-<framework>-<robot>-<task>-v0 naming convention.
• Checkpoints are provided for reproducibility and evaluation.
• Dynamic handover play supports command-file throw direction overrides: right_throw and left_throw.

License

This repository is released under the MIT License. See LICENSE.

Some files retain upstream Isaac Lab copyright and SPDX headers and remain subject to their original notices. See THIRD_PARTY_NOTICES.md.