RT-1
The original Robotics Transformer: 35M parameters, 3 Hz closed-loop control, trained on 130k teleoperated episodes across 700+ kitchen tasks.
- 97%
- 29K
- 3 Hz
- 130K
Benchmarks
Numbers as reported by the authors on the named evaluation suite.
| Evaluation | Suite | Score |
|---|---|---|
Seen tasks | RT-1 eval (700 tasks) | 97% |
Unseen tasks | RT-1 eval | 76% |
Distractor robustness | RT-1 eval | 83% |
Long-horizon SayCan | Kitchen SayCan | 67% |
Training
EfficientNet-B3 + TokenLearner + Transformer
35M
64x TPU v3
5.1K accelerator-hours
2.9K h
130K
Behaviour cloning with token-level cross-entropy over 256 action bins; TokenLearner compresses to 8 tokens/frame for 3 Hz inference.
Adding Kuka bin-picking and sim data raised unseen-task success without hurting seen tasks.
Linked datasets, robots & papers
Resolving related models, datasets and papers…