GEN-1

Generalist AI · 2026-04

VLANew in 2026Not disclosedNot disclosed

Pretraining data contain no robot data; trained on over 500,000 h of human wearable-device data.

Paper

No paper

Backbone LLM: Not disclosed
Vision encoder: —
Action head: Not disclosed (Not disclosed)
Chunk: Not disclosed
Control rate: Not disclosed
Architecture (original summary): Not disclosed (a large multimodal model that outputs actions in real time)

Pretraining accelerators: Not disclosed
Pretraining duration: Not disclosed
Pretraining steps & batch: Not disclosed
GPU/TPU-hours: Not disclosed
Fine-tuning: Hardware not disclosed; about 1 hour of robot data per task; in some tests matches GEN-0 with 10× less task data and fine-tuning steps
Data type: Human wearable-device data (physical interaction); no robot data in pretraining
Data scale: Over 500,000 hours (over half a million hours)
Main datasets: In-house dataset
Tasks & evaluation: Mean success 99% (GEN-0 fine-tuned 64%, no pretraining 19%); per task GEN-1 99% vs GEN-0 50%; auto-part kitting, T-shirt folding, box folding (summary tool values)
Robot: Not disclosed

Reinforcement learning: Used (VLA RL post-training) · Official blog states 'learning from experience (RL)' along with pretraining·post-training improvements. Applied stage·algorithm not disclosed
RL evidence: https://generalistai.com/blog/apr-02-2026-GEN-1
Base control: Fixed · No mention of base·locomotion (robot hand manipulation, model Not stated)

Key features: Pretraining data contain no robot data; trained on over 500,000 h of human wearable-device data. Reported 99% average success rate with about 1 h of robot data per task.
Notes: Posted 2026-04-02. Per-task figures (box folding 2.8× faster, etc.) are summarizer-tool values not yet checked against the original. Whether language is an input is not disclosed (note classifies it as VLA-class).
Verification: Confirmed