VLA Gallery / All models / Humanoid whole-body controllers

TWIST2

Unverified (Not stated in notes) · 2025-11

VAHumanoid whole-body controllersUnverifiedNot disclosed

Collects teleoperation data without mocap using PICO4U VR and a custom 2-DoF robot neck (about $250).

Paper

Backbone LLM: None
Vision encoder: —
Action head: Hierarchical visuomotor policy (Not disclosed)
Chunk: Not disclosed
Control rate: Not disclosed
Architecture (original summary): Unverified (architecture not extracted in notes); whole-body autonomous control via an egocentric-vision-based hierarchical visuomotor policy

Pretraining accelerators: Unverified (not extracted in notes)
Pretraining duration: Unverified (not extracted in notes)
Pretraining steps & batch: Unverified (not extracted in notes)
GPU/TPU-hours: Not disclosed
Fine-tuning: Unverified (not extracted in notes)
Data type: Real-robot teleoperation
Data scale: Unverified (not extracted in notes)
Main datasets: In-house collected dataset (public)
Tasks & evaluation: 100 demos collected in 15 minutes, near-100% success rate; egocentric vision-based autonomous whole-body control
Robot: Unverified (Not stated in notes)

Reinforcement learning: Not disclosed · Architecture·training method Unverified
Base control: Humanoid legged (Commands/latents → RL locomotion controller) · Autonomous diffusion policy outputs root motion and whole-body joint commands → low-level tracker executes joints including legs (kick, single step; loco-manipulation via teleoperation only)

Key features: Collects teleoperation data without mocap using PICO4U VR and a custom 2-DoF robot neck (about $250). Collects 100 demonstrations in 15 minutes with nearly 100% success.
Notes: Notes record only the arXiv abstract level (architecture·compute not extracted). System and dataset open-sourced; whether weights are released Unverified. Not covered by verification files.
Verification: Not verified