HPT-XLarge
Learns a shared policy trunk across 52 heterogeneous datasets by tokenizing each robot proprioception and vision stream into a common latent, then attaching lightweight task heads.
- 72%
- 19K
- 10 Hz
- 200K
Benchmarks
Numbers as reported by the authors on the named evaluation suite.
| Evaluation | Suite | Score |
|---|---|---|
Transfer to unseen embodiment | HPT eval | 20 |
Simulation average | HPT sim suite | 72% |
Real-world Franka tasks | HPT real | 66% |
Training
Heterogeneous pre-trained transformer trunk
1.1B
32x V100/A100
6.5K accelerator-hours
4.2K h
200K
Scaling study: trunk pre-training with embodiment-specific stems, then frozen-trunk transfer to new robots.
Frozen trunk + 20-demo head training beats full from-scratch training in low-data regimes.
Linked datasets, robots & papers
Resolving related models, datasets and papers…