Robot AI model
Heterogeneous Pre-trained Transformers
HPT pretrains a shared transformer trunk with embodiment-specific input stems and action heads; the published HPT-Base checkpoint is 12.6M parameters. Repository code uses MIT, while separately stated checkpoint license terms were not established.
Records last verified Oct 13, 2026 · Send a correction
Also known as HPT
- Organization
- MIT CSAIL / Meta FAIR
- Type
- Foundation model
- Released / introduced
- Sep 30, 2024
- Public weights
- Public weights
- Licence / scope
- Not established
- Parameters
- 12.6M
- Base model
- Not disclosed
- Humanoid association
- None reported
- Inputs
- vision, proprioception
- Outputs
- shared policy representations, robot actions through task-specific heads
Release dates can mark an announcement or paper introduction. Public checkpoints, code licences and weight terms are distinct; follow the publisher’s links for the stated scope.
Disclosed training data
- Heterogeneous mixture of robot teleoperation, simulation, deployed-robot data and human video
More robot AI models
Evidence
Sources
- [s1] Scaling Proprioceptive-Visual Learning with Heterogeneous Pre-trained Transformers ↗
MIT CSAIL / Meta FAIR · paper · published Sep 30, 2024 · accessed Oct 13, 2026
- [s2] HPT official project page ↗
MIT CSAIL / Meta FAIR · official · accessed Oct 13, 2026
- [s3] HPT-Base checkpoint model card ↗
Lirui Wang · official · accessed Oct 13, 2026
- [s4] HPT official code repository and MIT license ↗
HPT team · official · accessed Oct 13, 2026