salik/dex-arm-01-act

DEX ARM 01 — Biomimetic Dexterous Manipulator ACT Policy

Action Chunking with Transformers (ACT) imitation learning policy for DEX ARM 01, a 20-DoF biomimetic dexterous hand and 7-axis arm platform (the manipulation subsystem of the upcoming RAMA humanoid architecture).

• Interactive 3D WebGL / MuJoCo Simulation: salikshah.com/work/dex-arm-01/
• Demonstration Dataset: salik/dex-arm-01-sim
• Framework: Hugging Face LeRobot

---

Technical Specifications

Parameter	Specification
Platform	DEX ARM 01 (RAMA Subsystem)
Architecture	Action Chunking with Transformers (ACT)
Backbone	Transformer Encoder-Decoder with CVAE latent space
Chunk Size	100 action timesteps per prediction
State Dimension	27 (20-DoF hand joints + 7-axis arm joint states)
Action Dimension	27 (Target joint position commands)
Training Steps	20,000 steps on Tesla T4 GPU
Base Dataset	salik/dex-arm-01-sim (37 episodes, 1,722 frames @ 30 FPS)
Hardware Target	5-digit anthropomorphic hand (20 active DoF) + 7-axis arm

---

Quickstart / Inference

import torch
from lerobot.common.policies.act.modeling_act import ACTPolicy

# Download and instantiate DEX ARM 01 ACT policy
policy = ACTPolicy.from_pretrained("salik/dex-arm-01-act")
policy.eval()

# Example state input: [batch, state_dim]
batch_state = torch.zeros((1, 27))
with torch.no_grad():
    action_chunk = policy.select_action({"observation.state": batch_state})

print("Predicted action chunk shape:", action_chunk.shape)

---

Training Details

• Optimizer: AdamW (learning rate = 1e-5, weight decay = 1e-4)
• Loss Function: L1 position loss + KL-divergence weight beta = 10.0
• Temporal Ensembling: Exponential weighting across overlapping chunk predictions (tau = 0.01)
• Environment: DeepMind MuJoCo 3.x physics simulation

---

Attribution & Provenance

• Original Mechanical Design Reference: @crystalsssup on X
• AI Reasoning & Kinematics Prompting: Kimi 3.1 Pro
• Physics Engine: DeepMind MuJoCo 3.x (WASM runtime)
• Robotics Framework: Hugging Face LeRobot ACT
• Engineering & Policy Training: Salik Shah (salikshah.com)

---

License

Apache 2.0. Open for research, simulation, and hardware deployment.