GEN-1.5

Generalist AI · 2026-08

VLANew in 2026Not disclosedNot disclosed

Takes a 3–12 second demonstration as a physical prompt and performs new tasks without weight updates.

Paper

No paper

Backbone LLM: Not disclosed
Vision encoder: —
Action head: 100 Hz action trajectories (physical prompt) (Not disclosed)
Chunk: Not disclosed
Control rate: 100 Hz
Architecture (original summary): Not disclosed; input video (30 s context window)·sensors·language·proprioception, output 100 Hz action trajectories; in-context learning from sensorimotor 'physical prompts'

Pretraining accelerators: Not disclosed
Pretraining duration: Continuous pretraining for over 8 months (over 8 months)
Pretraining steps & batch: Not disclosed
GPU/TPU-hours: Not disclosed
Fine-tuning: Hardware not disclosed; few-shot adaptation in 1–10 gradient steps, 1–5 minutes of data (about 10–50 demonstrations), weight change under 0.15%
Data type: Not disclosed
Data scale: Not disclosed
Main datasets: Not disclosed
Tasks & evaluation: 10 tasks one-shot ICL average 59% (±10%), few-shot 83% (±9%, 10 gradient steps)
Robot: Not disclosed

Reinforcement learning: Not disclosed · Training method details not disclosed
Base control: Fixed · No mention of base·locomotion (bimanual manipulation, 100 Hz action trajectories)

Key features: Takes a 3–12 second demonstration as a physical prompt and performs new tasks without weight updates. One-shot 59%, few-shot with 10 gradient steps 83%.
Notes: Posted 2026-08-19. Few-shot data of 1–5 minutes · 1–10 steps are note values; verification confirmed '10 gradient steps'.
Verification: Confirmed