ACT V5: SO-101 Precision Pick-and-Place Baseline

Task

Pick up a red cube and put it in a box.

Hardware

  • SO-101 leader-follower robot arm
  • External RGB camera
  • NVIDIA RTX 5060 Laptop GPU

Training data

20 high-quality teleoperated demonstrations. The cube starts at a fixed reference location A.

Training configuration

  • Policy: ACT
  • Training steps: 20,000
  • Batch size: 8
  • Input: RGB image and robot joint state
  • Output: SO-101 joint action chunk

Evaluation

The policy was evaluated in 10 independent autonomous real-robot trials.

Metric Result
Successes 8 / 10
Success rate 80%

Result

This is the initial closed-loop imitation-learning baseline. The next stage adds a wrist camera, multiple cube positions, language instructions, and SmolVLA fine-tuning.

Limitations

The evaluation uses a fixed object position and a single external RGB camera. It does not yet test position, lighting, or language generalization.

Downloads last month
20
Safetensors
Model size
51.7M params
Tensor type
F32
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support