RevoLab

Isaac Lab environments and learning code for BrainCo Revo3 dexterous hands and Revotron arm-hand systems. Robot descriptions are supplied by the pinned brainco-description submodule.

The current task suite is v1, using the 20261014 asset release. This training campaign is complete: all six tasks have successful simulation demos below, and all campaign training has stopped. Completion means a successful demonstration was recorded for each task; it does not require every rollout to succeed. No v1 pretrained weights or validated hardware deployment results are published yet.

Tasks

Task ID	Training entry point	Robot asset	Policy status
BrainCo-Direct-Revo3-Repose-Cube-v1	RSL-RL	Revo3 right hand	Successful demo recorded; training stopped
BrainCo-Direct-Revo3-Reorient-Cylinder-v1	RSL-RL	Revo3 right hand	Successful demo recorded; training stopped
BrainCo-Dexsuite-Revo3-Right-Lift-v1	RSL-RL	Revotron single right arm + Revo3	Successful demo recorded; training stopped
BrainCo-Direct-Revo3-HoraRotate-Ball-v1	HORA PPO / ProprioAdapt	Revo3 right hand	Stage 1 complete; no further training planned in this campaign
BrainCo-Direct-Revo3-HoraRotate-Cylinder-v1	HORA PPO / ProprioAdapt	Revo3 right hand	Stage 1 complete; no further training planned in this campaign
BrainCo-Dynamic-Handover-Revo3-Cube-v1	RL-Games	Revotron bimanual + Revo3	Training complete; right-to-left transfer demonstrated

Lift and handover also provide -Play-v1 configurations. Handover locks the waist, neck and head at zero; actions control the 56 arm and finger joints.

Simulation demos

Training status and evaluation below are from 2026-10-19. Lift, Repose and Reorient were stopped after their demos were recorded; checkpoints and training logs are retained. The GIFs show selected successful simulation episodes or excerpts, not hardware results.

Lift	Repose · random cube orientations	Reorient · successive cylinder orientations
4 s; iteration 6600	8 s excerpt; iteration 4925; 14 targets in the full 10 s rollout	8 s excerpt; iteration 7125; 229 targets in the full 10 s rollout

These three clips show normal-speed rollouts from checkpoints collected during training. Lift brings the object to the blue position target; the original small vibrations remain visible. Repose's floating cube shows the current target orientation. Reorient advances the target yaw by 22.5° after each success; color blocks are added for display so its rapid rotation is visible. Robot and object motion is replayed directly from recorded simulation states, without interpolation or motion editing.

HORA Ball · Stage 1	HORA Cylinder · Stage 1	Handover · right to left
8 s excerpt at normal speed; about 2 turns in the full 20 s rollout	8 s excerpt at normal speed; about 2.5 turns in the full 20 s rollout	1.25 s episode, shown at 0.25× speed

The ball and cylinder use color blocks to make rotation visible. Only their visual surfaces change: the original analytic colliders, mass, inertia and friction are preserved. The animations replay recorded simulation states with the new appearance. Handover transfers the orange cube; the blue sphere marks its target.

For the checkpoints shown, a fixed-seed evaluation with 128 environments per condition gave:

Policy	Checkpoint epoch	Evaluation result
HORA Ball	8972	47/128 (36.7%) held for 20 s and rotated at least one turn
HORA Cylinder	1272	124/128 (96.9%) held for 20 s and rotated at least one turn
Handover	3157	Right to left: 93/128 (72.7%); left to right: 0/128

HORA is the learning pipeline adapted from In-Hand Object Rotation via Rapid Motor Adaptation, used here for the ball and cylinder rotation tasks. Stage 1 trains a PPO teacher with privileged simulation observations. Stage 2 trains an adaptation module to infer the teacher's latent physical information from observation history; it is a second training stage for the same tasks.

The HORA tasks use normal gravity (9.81 m/s²), grasp caches and domain randomization. Rotation is signed cumulative rotation along the target axis; holding requires no height termination during the 20 s episode. Both demonstrated policies are Stage 1 teachers; Stage 2 adaptation has not been trained and is outside this campaign's scope. No further training is planned in this campaign.

Handover success here requires release by the source hand, proximity to the receiving palm and target, low object speed, and at least two nearby fingertips, continuously for at least 0.20 s through the episode end. This is a geometric reception check over a short episode, not a contact-force or long-duration stability test. Its reset is nearly fixed, so the 128 environments do not represent 128 independent starting scenes. The current policy has not learned reliable bidirectional handover; the equal-direction aggregate is 93/256 (36.3%).

Setup

The target environment is Isaac Sim 5.1, Python 3.11 and the pinned Isaac Lab commit documented in remote environment setup, on Linux with an NVIDIA GPU. The tested checkout's VERSION file says 2.3.2, but it is newer than that release tag. This repository is an editable Isaac Lab extension; a standalone wheel does not include robot assets.

git clone https://github.com/BrainCoTech/RevoLab.git
cd RevoLab
# Install Git LFS using your OS package manager first.
bash scripts/setup_assets.sh
# Run with the Python environment configured for Isaac Lab:
python -m pip install -e source/BrainCo_DexHand

setup_assets.sh checks out the gitlink recorded by RevoLab and fetches LFS objects. Updating the submodule to a newer upstream commit is a deliberate code change; do not use git submodule update --remote for routine setup.

Validation without training

Offline checks require only the USD Python bindings:

python -m pip install usd-core
python scripts/check_assets.py

In the Isaac Lab Python environment, reset and step a task with zero actions:

python scripts/smoke_tasks.py --task BrainCo-Direct-Revo3-Repose-Cube-v1 --num_envs 2 --steps 32 --headless

Use any task ID from the table. The check validates finite outputs, default joint limits, named palm frames, and the handover joint locks. It does not measure task success or policy performance. CI runs the offline check; GPU checks run separately in Isaac Sim. See the local validation record for the tested build and results.

The colored HORA objects also have a GPU comparison against their plain counterparts:

python scripts/check_rotation_visuals.py --headless

This checks cloned visuals, unchanged physics attributes, one collider per object, matching mass/inertia/materials, and matching motion through 120 fall-and-contact steps.

Remote environments and all six tasks have also been checked on three GPU servers; see the remote validation record and server setup and launch commands.

Learning entry points

Use the Isaac Lab Python environment for training:

python scripts/rsl_rl/train.py --task BrainCo-Direct-Revo3-Repose-Cube-v1 --headless
python scripts/rsl_rl/train.py --task BrainCo-Direct-Revo3-Reorient-Cylinder-v1 --headless
python scripts/rsl_rl/train.py --task BrainCo-Dexsuite-Revo3-Right-Lift-v1 --headless
python scripts/rl_games/train.py --task BrainCo-Dynamic-Handover-Revo3-Cube-v1 --headless
python scripts/hora/train.py --algo PPO --task ball --cache_file /path/to/new-ball-cache.npz --headless
python scripts/hora/train.py --algo PPO --task cylinder --cache_file /path/to/new-cylinder-cache.npz --headless
python scripts/hora/train.py --algo ProprioAdapt --task ball --checkpoint /path/to/stage1_nn/best.pth --headless

Each new run writes params/asset_contract.json with the asset commit, ordered joints, limits and observation/action spaces. Resume, evaluation and export require matching metadata. Keep this file together with checkpoints when moving a run. See asset conventions for HORA cache format and migration details.

Layout

third_party/brainco-description/  # Pinned upstream robot descriptions (submodule)
assets/                         # Task objects and ground only
source/BrainCo_DexHand/          # Isaac Lab tasks, asset adapters and HORA algorithms
scripts/                        # Setup, validation, training and export
checkpoints/                    # Artifact status; generated weights are ignored
deploy/revo3/                   # ONNX deployment tooling; new calibration is pending
docs/                           # Asset contracts and migration validation
tests/                          # Offline regression checks
LICENSES/                       # Upstream license texts

See the documentation index, changelog and contribution guide.

RevoLab code is MIT licensed, with third-party code and assets described in THIRD_PARTY_NOTICES.md. Issues and contributions are welcome.