OpenHLM

Tsinghua + Shanghai Qi Zhi Institute + Spirit AI · 2026-06

VLANew in 2026Inherited from π0.5flow matching

A humanoid recipe that fine-tunes π0.5 to output whole-body commands and stacks it on the SONIC low-level whole-body controller.

Paper

Backbone LLM: Gemma 2B · PaliGemma 3B config (π0.5)
Vision encoder: SigLIP 400M (π0.5)
Action head: flow matching head (10 steps) (flow matching)
Chunk: 50 (25 executed)
Control rate: 10 Hz (SONIC 50 Hz)
Architecture (original summary): π0.5 (PaliGemma 3B) + flow matching action head (10 denoising steps); 32-D whole-body commands + 2-D gripper = 34-D output; SONIC tracker 50 Hz on top of VLA 10 Hz, executes 25 actions per inference, action horizon 50

Pretraining accelerators: N/A (initialized from π0.5)
Pretraining duration: N/A
Pretraining steps & batch: N/A
GPU/TPU-hours: Not disclosed
Fine-tuning: 4 × A800, about 24 hours, 30k steps, batch 128
Data type: Real-robot teleoperation (whole-body) + human demonstrations (HuMI robot-free demonstrations)
Data scale: 320 teleoperation demonstrations (8 tasks × 40) + HuMI robot-free demonstrations + 120 long-horizon demonstrations
Main datasets: In-house HLM-12 demonstrations, HuMI
Tasks & evaluation: HLM-12 (Unitree G1): average progress 89% on 8 training tasks; unseen tasks 33% without co-training·87% with; long-horizon fruit sorting 87.5% (1.14 hours of demos; GR00T N1.6 57.5%, Ψ0 48.8%, each 2.70 hours)
Robot: Unitree G1

Reinforcement learning: Partial (RL low-level controller) · π0.5 fine-tuning (supervised learning); whole-body execution by SONIC tracker trained with RL
Base control: Humanoid legged (Commands/latents → RL locomotion controller) · VLA outputs chunks of 32-D whole-body commands (root roll·pitch·yaw angular velocity + 29 joint targets including legs) → executed by SONIC tracker (walking to a table·pushing a cart)

Key features: A humanoid recipe that fine-tunes π0.5 to output whole-body commands and stacks it on the SONIC low-level whole-body controller. Co-training raised progress on unseen tasks from 33% to 87%.
Notes: A fine-tuning recipe, not a new pretrained model (note classification). arXiv 2026-06-20. Code·training data·checkpoints released. No sentence on total parameters, so numeric columns left empty. No matching row in the verification note.
Verification: Not verified