EgoScale

NVIDIA GEAR + UC Berkeley + University of Maryland · 2026-02

VLANew in 2026Not disclosedflow matching

Pretrains a VLA on over 20,854 hours of egocentric human video, raising average success on dexterous tasks by 54%.

Paper

Backbone LLM: Not stated (VL backbone of GR00T N1 design)
Vision encoder: —
Action head: DiT flow expert + per-embodiment MLP (flow matching)
Chunk: Not disclosed
Control rate: Not disclosed
Architecture (original summary): GR00T N1 design: VL backbone + DiT action expert (flow) + per-embodiment MLP adapters; relative SE(3) wrists + 22-DoF hand joints (retargeted from 21 human hand keypoints)

Pretraining accelerators: 256 × GB200 (Stage I human data pretraining)
Pretraining duration: Not disclosed
Pretraining steps & batch: Stage I 100K steps, batch 8,192, LR 5×10⁻⁵
GPU/TPU-hours: Not disclosed
Fine-tuning: Stage II (human-robot alignment mid-training) 50K steps·batch 2,048, Stage III (post-training) 10K steps·batch 512; GPU·time not stated
Data type: Human video (egocentric, action-labeled) + real-robot teleoperation (alignment mid-training)
Data scale: Over 20,854 hours of egocentric human video (incl. 829 hours of EgoDex); mid-training about 50 hours human + 4 hours robot (344 tasks)
Main datasets: In-house in-the-wild egocentric recordings, EgoDex
Tasks & evaluation: 5 dexterity tasks (shirt rolling·card sorting·moving fruit with tongs·bottle cap·syringe); average success rate +54% vs no pretraining; one-shot shirt folding 0.88·water bottle opening 0.55
Robot: Galaxea R1Pro + Sharpa Wave 22-DoF hands; Unitree G1 + 7-DoF three-finger hands (lower body HOMIE)

Reinforcement learning: Not used · Human video pretraining → alignment mid-training → demonstration fine-tuning (supervised)
RL evidence: https://arxiv.org/html/2602.16710
Base control: Fixed · R1 Pro base·torso explicitly fixed (arm relative EE + hands); G1 outputs only upper-body targets, lower-body balance·walking by a separate HOMIE policy

Key features: Pretrains a VLA on over 20,854 hours of egocentric human video, raising average success on dexterous tasks by 54%. Average completion rose from 0.30 to 0.71 as data grew from 1k to 20k hours.
Notes: Corrections applied: 256 GB200 is a sentence tied only to Stage I; Stage II·III GPUs not stated. Scenes 9,869·tasks 6,015·objects 43,237 refer to the in-the-wild recordings. Conflicts with the project page's '22k Hours' label, so the paper value is used. Basis of GR00T N1.7's human data. No primary evidence of weights·data release (unverifiable). Data hours (h) are the Stage I human video portion. Validation loss L = 0.024 − 0.003·ln(D), R² 0.9983.
Verification: Corrections applied