gr00t-n1.5
Candidate3B params
Overview
V1.x candidate — filename now confirmed, launch args verified
Architecture
- Vision backbone
- SigLIP2-400M (Eagle-2.5 VLM)
- Language backbone
- Qwen3-1.7B
- Action head
- AltVL DiT + self-attention (diffusion)
- Action steps per chunk
- 16
- Solver steps
- 4
- Camera views
- 2 — video.image, video.wrist_image
- Hardcoded keys, different naming convention from observation.images.* used by other archs.
Hardware fit
Fits the 8GB Orin Nano target
split only — simulator offloaded to a second machine, benchmark includes network hop
Benchmarks
| Hardware | Success rate | Step time | Inference time | Memory |
|---|---|---|---|---|
| RTX 3060 | 96% | 14.17 ms | 147 ms | 4866 MiB VRAM |
| Orin Nano 8GB (split) | — | 84.76 ms | — | — |