gr00t-n1.6

Overview

~6GB resident weights exhaust the 8GB unified memory pool entirely — does not fit even with the simulator offloaded elsewhere. Only usable on higher-tier hardware (AGX Orin or better). Used in the paper's real-robot ALOHA stress test (87.5% success vs 40% for PyTorch baseline) — good architecture, wrong hardware tier for this product.

Architecture

Vision backbone

SigLIP2-400M (Eagle3-VL)

Language backbone

Qwen3-1.7B

Action head

AltVL DiT (diffusion)

Action steps per chunk

16

Solver steps

4

Camera views

ambiguous

3 mutually-incompatible real deployments exist in this repo: LIBERO (2 views: video.image, video.wrist_image), SimplerEnv-bridge (1 view: image_0), ALOHA (1 view, name unconfirmed).

Hardware fit