Ψ0 (Psi-Zero)

USC PSI Lab + NVIDIA + WorldEngine · 2026-03

VLANew in 2026Over 2.5Bhybrid

Reported beating baselines trained on over 10× more data by 40%+ using only about 800 h of human video and 30 h of robot data.

Paper

Backbone LLM: Qwen3 · Qwen3-VL-2B config
Vision encoder: Qwen3-VL vision
Action head: MM-DiT action expert about 500M (hybrid)
Chunk: Not disclosed
Control rate: Not disclosed
Architecture (original summary): Qwen3-VL-2B-Instruct + MM-DiT action expert (about 500M); pretraining on FAST tokens (48→about 20) for 48-DoF task-space actions, post-training on 36-DoF joint-space actions; lower body uses RL tracking policy AMO (8-DoF commands→15-DoF joints)

Pretraining accelerators: 64 × A100
Pretraining duration: 10 days
Pretraining steps & batch: Global batch 1024, LR 0.0001 constant (steps Not disclosed)
GPU/TPU-hours: 15,360 (Converted · A100-hours (64 GPUs × 10 days = 240 hours))
Fine-tuning: Post-training on 32 A100s (single node) about 30 h, batch 2048; per-task fine-tuning 40,000 steps (cosine LR starting at 1e-4)
Data type: Human video (egocentric EgoDex) + real-robot teleoperation (humanoid)
Data scale: EgoDex about 829 hours; Humanoid Everyday 31 hours (260 tasks); 80 teleoperation trajectories per task
Main datasets: EgoDex, Humanoid Everyday
Tasks & evaluation: Unitree G1 8 real tasks; overall success rate 40%+ higher than 2nd-place GR00T-N1.6; compared π0.5, GR00T N1.6, EgoVLA, H-RDT, InternVLA-M1, DP, ACT
Robot: Unitree G1 (29 DoF, Dex3-1 7-DoF hands, head RealSense D435i)

Inference device & speed: Single forward pass about 160 ms (hardware Not stated)
Inference memory: Not disclosed
Weights bf16 (calc.): 5 GB
Where it runs: Not disclosed

Reinforcement learning: Partial (RL low-level controller) · VLA by supervised learning; lower body by RL tracking policy AMO (8-DoF commands → 15-DoF joints)
Base control: Humanoid legged (Commands/latents → RL locomotion controller) · VLA outputs hand·arm joints + lower-body 8-DoF commands (torso rpy·height·vx·vy·vyaw·direction) → AMO RL tracking policy generates 15-DoF leg·waist joints (turning·walking·squatting)

Key features: Reported beating baselines trained on over 10× more data by 40%+ using only about 800 h of human video and 30 h of robot data. Handles real-time action chunking at training time.
Notes: Corrections applied: weight release confirmed (HF USC-PSI-Lab/psi-model, data psi-data, Apache 2.0); checkpoint for SONIC added 2026-09-13. arXiv 2026-03-12. Best Paper Award at the CVPR 2026 3D-LLM/VLA workshop (2026-06-03). Parameters (B) is the lower bound 2.5. Data hours (h) are the pretraining EgoDex portion.
Verification: Corrections applied