Overview
ML6’s Robotics & AI team conducted practical experiments with imitation learning using LeRobot, testing both narrow (ACT) and foundation models (GR00T-N1) on custom datasets with SO-ARM100 robots.
Narrow Models (ACT)
Performance results:
- Simple single-position task: 60% accuracy (10k frames)
- Five-position vertical axis: 90% accuracy (46k frames)
- Two-dimensional grid: 79% accuracy (137k frames)
Limitations: minimal generalization beyond training distribution; failed with out-of-distribution camera variations.
Foundation Models (GR00T-N1)
Performance results:
- Textile spreading: 60% accuracy (53k frames, 29 min recording)
- Towel folding: 80% accuracy (76k frames, 42 min recording)
Limitations: stuttering motion from inference latency; struggles with precision in high-variability tasks.
Critical Success Factors for Datasets
- Accuracy — carefully recorded data is essential
- Controlled sequences — simplified, sequential movements
- Comprehensive coverage — exposure to all task variations
- Robustness — include error recovery scenarios
Industry Assessment
Imitation learning is “closer than most expect” for production robotics in controlled environments with repetitive, structured tasks. GR00T-N1.5 + asynchronous inference + NVIDIA Jetson Thor are rapidly addressing latency and compute constraints.
ML6 placed 3rd in the 2025 LeRobot Hackathon using Gaussian splatting for camera instability.