BrainCo Isaac Lab Simulation

Overview

BrainCoTech/RevoLab is an official extension package built on NVIDIA Isaac Lab. It provides BrainCo dexterous-hand robot assets, reinforcement-learning environments, pretrained checkpoints, and sim-to-real deployment tools. It is not a standalone Isaac Lab fork, so a working Isaac Lab installation is required.

Requirements

• Python 3.10 or later
• A working NVIDIA Isaac Lab installation

Installation

Clone the repository and install the BrainCo extension in the active Isaac Lab Python environment:

git clone https://github.com/BrainCoTech/RevoLab.git
cd RevoLab/source/BrainCo_DexHand
pip install -e .

Verify that the task extension can be imported:

python -c "import BrainCo_DexHand"

Included Tasks

Robot	Framework	Task ID	Description
Revo 3	Direct	BrainCo-Direct-Revo3-Repose-Cube-v0	In-hand cube repose
Revo 3	Direct	BrainCo-Direct-Revo3-Reorient-Cylinder-v0	In-hand cylinder reorientation
Revo 3	Dexsuite	BrainCo-Dexsuite-Revo3-Right-Lift-v0	Right-hand grasp and lift
Revo 3	HORA	BrainCo-Direct-Revo3-HoraRotate-Ball-v0	In-hand ball rotation
Revo 3	HORA	BrainCo-Direct-Revo3-HoraRotate-Cylinder-v0	In-hand cylinder rotation
RevoTron	Dynamic Handover	BrainCo-Dynamic-Handover-Revo3-Cube-v0	Dynamic two-hand cube handover

Downloading Pretrained Checkpoints

Run the download script from the repository root to populate checkpoints/:

./scripts/download-checkpoints.sh

Training and Evaluation

The Direct and Dexsuite tasks use RSL-RL. For example, train and evaluate the cube-repose task with:

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

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

The HORA rotation tasks use a separate PPO / ProprioAdapt training path and checkpoint format. The RevoTron dynamic-handover task uses RL-Games. See the RevoLab repository README for their training, evaluation, and ONNX export commands.

Sim-to-real Deployment

deploy/revo3 provides a lightweight, ROS-free package for running exported ONNX policies on a physical Revo 3 hand through the Revo 3 Python SDK. The runtime reads measured joint positions, builds policy observations, performs ONNX inference, and sends closed-loop MIT commands to the hand.

cd deploy/revo3
pip install -e .
pip install -e ".[hardware]"

Before deployment, configure policy.onnx, policy.yaml, and config/revo3_right.yaml for the target hardware. See deploy/revo3/README.md for complete export, dry-run, and hardware execution instructions.