Action Chunking with Transformers (ACT) is an imitation-learning method that predicts short action chunks instead of single steps. It learns from teleoperated data and often achieves high success rates. This policy has been trained and pushed to the Hub using LeRobot. Learn how to train and run it in the LeRobot act guide, or browse the full documentation. The policy consumes these observation features and produces these action features. Inputs Outputs New to LeRobot? These guides cover the full workflow: - Install LeRobot — set up the lerobot package. - Hardware setup — assemble, wire, and calibrate your robot and cameras. - Record data & train a policy — the end-to-end imitation-learning…
Open-weight model · Robotics
soarm_amazing_hand_act
by Juxi Technology Juxi-Technology/soarm_amazing_hand_act
An ACT (Action Chunking Transformer) imitation-learning policy trained on an SO-ARM101 follower arm equipped with an AmazingHand dexterous hand, performing a cube pick-up task.
Parameters52M
Context—
Weights206.7 MB
Licenseapache-2.0
AccessOpen weights
Monthly Downloads—
Runs On
What it takes to serve soarm_amazing_hand_act (52M parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.
| Precision | Weights | Memory needed | Cheapest setup | Per hour | Also fits |
|---|---|---|---|---|---|
| 16-bit | 0.1 GB | 0.1 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 8-bit | 0.1 GB | 0.1 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 4-bit | 0.0 GB | 0.0 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Sep 18, 2026.
Model Card
By Juxi Technology, published under apache-2.0, revision adf7e66fabbc.
An ACT (Action Chunking Transformer) imitation-learning policy trained on an SO-ARM101 follower arm equipped with an AmazingHand dexterous hand, performing a cube pick-up task. - 20 teleoperated demonstrations, 18,538 frames, 30 fps - The cube was recorded at 4 different table positions, 5 episodes each, to cover positional variation shoulderpan.pos, shoulderlift.pos, elbowflex.pos, wristflex.pos, wristroll.pos, gripper.pos Requires an environment matching this project (a customized lerobot that includes the soamazinghand robot definition): 1. --task must match the training singletask string exactly 2. Camera names and index order must match training 3. The robot must already be calibrated…
Read Juxi Technology's full model card
SO-ARM101 + AmazingHand — ACT Grasping Policy
An ACT (Action Chunking Transformer) imitation-learning policy trained on an SO-ARM101 follower arm equipped with an AmazingHand dexterous hand, performing a cube pick-up task.
Hardware
| Component | Model | Notes |
|---|---|---|
| Follower arm | SO-ARM101 | 5 × STS3215 (IDs 1–5); original gripper servo #6 removed |
| End-effector | AmazingHand | 8 × SCS0009 (IDs 1–8); dedicated serial port + separate power supply |
| Leader arm | SO-ARM101 | Full 6 servos; servo #6 (gripper) drives the hand's open/close |
| Cameras | 2 × USB camera | top (overhead) + wrist, 640×480 @ 30 fps |
Implementation notes: The hand is driven through
rustypot(Scs0009PyController) rather than lerobot's Feetech bus — SCS0009 (protocol 1) and STS3215 (protocol 0) cannot share a bus. The leader gripper position is mapped linearly onto the hand's open/close pose.
Training Data
- Dataset:
Juxi-Technology/soarm_amazing_hand_pick - 20 teleoperated demonstrations, 18,538 frames, 30 fps
- Task:
Pick up the cube with the dexterous hand - The cube was recorded at 4 different table positions, 5 episodes each, to cover positional variation
Training Details
| Setting | Value |
|---|---|
| Policy | ACT |
| Steps | 60,000 |
| Batch size | 8 |
| Optimizer | AdamW, lr 1e-5, weight decay 1e-4, grad clip 10.0 |
| Vision backbone | ResNet-18 (ImageNet pretrained) |
| chunk_size / n_action_steps | 100 / 100 |
| Inputs | observation.state (6) + 2 RGB images |
| Output | action (6) |
Joint order (6-dim):
shoulder_pan.pos, shoulder_lift.pos, elbow_flex.pos, wrist_flex.pos, wrist_roll.pos, gripper.pos
gripper.posis a virtual dimension — the leader gripper opening, mapped to the hand's grasp pose.
Usage
Requires an environment matching this project (a customized lerobot that includes the so_amazing_hand robot definition):
pip install -e ".[amazinghand,training]"
lerobot-rollout \
--strategy.type=base \
--policy.path=Juxi-Technology/soarm_amazing_hand_act \
--device=cuda \
--robot.type=so101_amazing_hand \
--robot.port=<follower-port> \
--robot.hand_port=<hand-port> \
--robot.id=amazing_hand_follower \
--robot.cameras='{
top: {type: opencv, index_or_path: <top-camera-index>, width: 640, height: 480, fps: 30},
wrist: {type: opencv, index_or_path: <wrist-camera-index>, width: 640, height: 480, fps: 30}
}' \
--task="Pick up the cube with the dexterous hand" \
--duration=60
Preconditions:
--taskmust match the trainingsingle_taskstring exactly- Camera names and index order must match training
- The robot must already be calibrated (
hand_angles.json, per-joint arm calibration files) - The hand must be on its own serial port with an independent power supply
Limitations
- Limited positional generalization: performs well near the 4 recorded cube positions; success rate drops noticeably elsewhere
- Single-shot policy: does not loop autonomously after completing a grasp — the object and the arm must be reset manually
- Environment sensitive: collected under a single scene and lighting condition; changes to camera pose, lighting, or background degrade performance
- GPU deployment required: CPU inference runs at ~4 Hz (against a 30 Hz target), making motion visibly slow — use an NVIDIA GPU (
--device=cuda)
Related
- LeRobot — training and deployment framework
- SO-ARM101 — open-source 6-DoF robotic arm
- AmazingHand — open-source dexterous hand
Identity and Version
- Repository
- Juxi-Technology/soarm_amazing_hand_act
- Publisher
- Juxi Technology
- Task
- Robotics
- Modality
- Control
- Library
- lerobot
- Parameters
- 52M parameters
- Languages
- act, so-101, so-arm101
- Revision
- adf7e66fabbcec5b3d99bbbec8d2b27dfc3fb7cc
- First published
- 2026-09-18
- Last updated
- 2026-09-18
Files and Weights
9 files, 206.7 MB in total. The weights are 3 files totalling 206.7 MB in safetensors.
Weights3 files · 206.7 MB
Configuration4 files · 11.2 KB
Documentation1 file · 4.0 KB
Repository1 file · 1.5 KB
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| model.safetensors | Weights | 206.7 MB | dc8f82ba902a |
| policy_postprocessor_step_0_unnormalizer_processor.safetensors | Weights | 7.5 KB | fb4fc4390c40 |
| policy_preprocessor_step_3_normalizer_processor.safetensors | Weights | 7.5 KB | a6125d603ffb |
| config.json | Configuration | 1.7 KB | — |
| policy_postprocessor.json | Configuration | 660 B | — |
| policy_preprocessor.json | Configuration | 1.3 KB | — |
| train_config.json | Configuration | 7.4 KB | — |
| README.md | Documentation | 4.0 KB | — |
| .gitattributes | Repository | 1.5 KB | — |
License and Download
- License
- apache-2.0
- Access
- Open weights, no gate
- Download size
- 206.7 MB
Download from Juxi Technology
Released by Juxi Technology through its official repository on Hugging Face. Read the license.
Built From
- Described by arXiv:2304.13705
- Trained on (disclosed) Juxi-Technology/soarm_amazing_hand_pick
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 206.7 MB |
| 16-bit | 0.1 GB |
| 8-bit | 0.1 GB |
| 4-bit | 0.0 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Questions About soarm_amazing_hand_act
How much GPU memory does soarm_amazing_hand_act need?
About 0.1 GB at 16-bit and 0 GB at 4-bit: the weights (52M parameters) plus a working margin. A long context needs more.
What is the cheapest GPU to run soarm_amazing_hand_act on?
At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.
Can I use soarm_amazing_hand_act commercially?
Yes. soarm_amazing_hand_act is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.
Similar Models
Action Chunking with Transformers (ACT) is an imitation-learning method that predicts short action chunks instead of single steps. It learns from teleoperated data and often achieves high success rates. This policy has been trained and pushed to the Hub using LeRobot. Learn how to train and run it in the LeRobot act guide, or browse the full documentation. The policy consumes these observation features and produces these action features. Inputs Outputs New to LeRobot? These guides cover the full workflow: - Install LeRobot — set up the lerobot package. - Hardware setup — assemble, wire, and calibrate your robot and cameras. - Record data & train a policy — the end-to-end imitation-learning…
Action Chunking with Transformers (ACT) is an imitation-learning method that predicts short action chunks instead of single steps. It learns from teleoperated data and often achieves high success rates. This policy has been trained and pushed to the Hub using LeRobot. Learn how to train and run it in the LeRobot act guide, or browse the full documentation. The policy consumes these observation features and produces these action features. Inputs Outputs New to LeRobot? These guides cover the full workflow: - Install LeRobot — set up the lerobot package. - Hardware setup — assemble, wire, and calibrate your robot and cameras. - Record data & train a policy — the end-to-end imitation-learning…
Action Chunking with Transformers (ACT) is an imitation-learning method that predicts short action chunks instead of single steps. It learns from teleoperated data and often achieves high success rates. This policy has been trained and pushed to the Hub using LeRobot. Learn how to train and run it in the LeRobot act guide, or browse the full documentation. The policy consumes these observation features and produces these action features. Inputs Outputs New to LeRobot? These guides cover the full workflow: - Install LeRobot — set up the lerobot package. - Hardware setup — assemble, wire, and calibrate your robot and cameras. - Record data & train a policy — the end-to-end imitation-learning…
Action Chunking Transformer Policy (as per Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware) trained for the AlohaTransferCube environment from gym-aloha. See the LeRobot library (particularly the evaluation script) for instructions on how to load and evaluate this model. Trained with LeRobot@3c0a209. The model was trained using LeRobot's training script and with the alohasimtransfercubehuman dataset, using this command: The training curves may be found at https://wandb.ai/aliberts/lerobot/runs/720l37xb. The current model corresponds to the checkpoint at 80k steps. This took about 1h45 to train on an Nvida A100. The model was evaluated on the AlohaTransferCube task from…
Action Chunking with Transformers (ACT) is an imitation-learning method that predicts short action chunks instead of single steps. It learns from teleoperated data and often achieves high success rates. This policy has been trained and pushed to the Hub using LeRobot. See the full documentation at LeRobot Docs. For a complete walkthrough, see the training guide. Below is the short version on how to train and run inference/eval: Writes checkpoints to outputs/train/ /checkpoints/. Prefix the dataset repo with eval\ and supply --policy.path pointing to a local or hub checkpoint.