luisd7271/stationary-WidgetAssembly-Lerobot-act-v3-model

stationary-WidgetAssembly-Lerobot-act-v3

Model Summary

This model is an Action Chunking Transformer (ACT) policy trained to perform a widget assembly task using the Trossen Robotics Stationary AI Kit. The task involves picking up two interlocking components and connecting them at the correct angle and position. The model achieved a 100% success rate across 20 evaluation trials with consistent piece placement.

Hardware

• Robot: Trossen Robotics WidowX AI — 2 leader arms + 2 follower arms (bimanual stationary config)
• Cameras: 4x Intel RealSense D405 (cam_high, cam_low, cam_left_wrist, cam_right_wrist) at 640x480 @ 30fps
• Training machine: System76, with Intel Core i9, NVIDIA GeForce RTX 4060 Laptop GPU (8GB VRAM)

Training

• Policy: ACT (Action Chunking Transformer)
• Framework: HuggingFace LeRobot
• Dataset: 25 episodes, 20-30 seconds per episode
• Batch size: 4
• Steps: 100,000
• Final loss: ~0.058
• Training time: ~9 hours

Dataset

The dataset was recorded using leader/follower teleoperation. Recording practices that led to successful training are:

• Tape markers on the work surface ensured consistent piece placement across all episodes
• Smooth, deliberate arm movements with no jerky or rushed motions
• Arms returned to a consistent starting position at the end of every episode
• Components were held and placed by the arms rather than pushed or nudged

Evaluation

Metric	Result
Success rate (standard conditions)	100% (20/20)
Success rate (color-swapped pieces)	~60-70%
Self-correction behavior	Observed

Notable Findings

Self-correction: The model demonstrated emergent self-correcting behavior — if an assembly attempt failed, the arms would readjust and retry until the task was completed. This behavior was not explicitly demonstrated during training and is likely a result of ACT's closed-loop chunk-based replanning.

Color generalization: When piece colors were swapped, the model still completed the task at a reduced success rate (~60-70%), suggesting the policy learned geometric and spatial features rather than purely color-based cues.

Usage

# Evaluation command
uv run lerobot-record \
    --robot.type=bi_widowxai_follower_robot \
    --robot.left_arm_ip_address=192.168.1.5 \
    --robot.right_arm_ip_address=192.168.1.4 \
    --robot.id=bimanual_follower \
    --policy.path=luisd7271/stationary-WidgetAssembly-Lerobot-act-v3-model \
    --robot.min_time_to_move_multiplier=10.0 \
    --robot.left_arm_max_relative_target=0.2 \
    --robot.right_arm_max_relative_target=0.2

Limitations

• Piece placement must be within the pre-defined tape marker boundaries for reliable performance
• Color changes to the pieces reduce success rate
• Task is specific to this widget geometry and may not generalize to other assembly tasks without retraining

Citation

Trained by Luis at Trossen Robotics, June 2026.