1X Redwood AI

1X Technologies · 2025-06

VLAHumanoid VLA160Mdiffusion

Predicts locomotion, manipulation, and pelvis pose commands together with arm and hand commands for whole-body, multi-contact manipulation.

Paper

No paper

Backbone LLM: None (160M policy)
Vision encoder: Pretrained ViT
Action head: Diffusion policy (diffusion)
Chunk: Not disclosed
Control rate: About 5 Hz
Architecture (original summary): Pretrained language embedding + pretrained ViT vision tokens + joint position·applied force proprioception embedding fused → diffusion policy decodes EVE·NEO actions, about 5 Hz

Pretraining accelerators: Not disclosed
Pretraining duration: Not disclosed
Pretraining steps & batch: Not disclosed
GPU/TPU-hours: Not disclosed
Fine-tuning: Not disclosed
Data type: Real-robot teleoperation
Data scale: Not disclosed (teleoperation·autonomous episodes)
Main datasets: EVE · NEO teleoperation and autonomous episodes (collected at 1X offices · employee homes)
Tasks & evaluation: Fetching items, opening doors, moving around the home; whole-body·multi-contact manipulation, mobile bimanual manipulation, bracing·leaning the body during manipulation; success rates Not disclosed
Robot: 1X NEO Gamma (humanoid); training data includes wheeled EVE

Reinforcement learning: Not disclosed · Training method not disclosed
Base control: Humanoid legged (Format not stated) · VLA predicts locomotion commands along with arm·hand commands (joint navigation·manipulation planning); no low-level locomotion controller mentioned

Key features: Predicts locomotion, manipulation, and pelvis pose commands together with arm and hand commands for whole-body, multi-contact manipulation. Learns from both successful and failed rollouts.
Notes: Blog posted 2025-06-10. Training compute·data scale were NOT FOUND in verification either.
Verification: Confirmed