← Leaderboard

HPT (Heterogeneous Pre-trained Transformers)

Policy
MIT CSAIL · Sep 30, 2024

Pretraining recipe from MIT CSAIL and Meta FAIR: a shared transformer trunk is pretrained on proprioception and vision from 52 datasets, with small embodiment-specific stems and heads, and scaled past 1B parameters. It targets transfer to new embodiments rather than language-conditioned control. Its SimplerEnv results cover the Google Robot tasks only. The MIT license is the code repo’s; the checkpoints carry no license tag.

Type
Policy
Parameters
—
Backbone
Shared transformer trunk + embodiment-specific stems and heads
Weights
Open
License
MIT
Commercial use
Unclear
Released
Sep 30, 2024
Robots
Franka Panda, Google Robot (sim)
SO-101
–
LeRobot
–
Inputs
RGB · P
VRAM
–

Benchmark results

0 of 13 benchmarks

No benchmark results recorded yet.

Not reported on 13 benchmarks
Wrong or missing number? Report an error