BRS (WB-VIMA)

Stanford University · 2025-03

VAVANot discloseddiffusion

Learns long-horizon household whole-body manipulation with Galaxea R1, which has a mobile base·4-DoF torso·two arms.

Paper

Backbone LLM: None
Vision encoder: PointNet point cloud
Action head: 3 diffusion UNet heads (base, torso, arm) (diffusion)
Chunk: Not disclosed
Control rate: 10 Hz (controller 100 Hz)
Architecture (original summary): PointNet point cloud tokens + proprioception → causal self-attention trunk; whole-body actions decoded autoregressively by 3 independent diffusion networks for base·torso·arm (lightweight UNet heads); data 10 Hz, controller 100 Hz

Pretraining accelerators: N/A
Pretraining duration: N/A
Pretraining steps & batch: N/A
GPU/TPU-hours: Not disclosed
Fine-tuning: Not disclosed (no GPU·time·steps·batch; README training command has a gpus=<NUM_GPUS> placeholder)
Data type: Real-robot teleoperation
Data scale: 100, 103, 98, 138, 122 trajectories per task
Main datasets: Self-collected with JoyLo
Tasks & evaluation: 5 long-horizon household tasks
Robot: Galaxea R1 (2 × 6-DoF arms + 4-DoF torso + omnidirectional mobile base)

Reinforcement learning: Not used · Imitation learning on whole-body teleoperation data
Base control: Wheeled (Direct base velocity/pose output) · WB-VIMA directly outputs omnidirectional base actions (3 DoF) via a dedicated diffusion head, autoregressively decoded with torso·arms

Key features: Learns long-horizon household whole-body manipulation with Galaxea R1, which has a mobile base·4-DoF torso·two arms. Collected 98–138 trajectories per task with the JoyLo teleoperation interface.
Notes: A borderline case for scope: whole-body mobile bimanual without legs. The UNet head description was not in the verification extraction.
Verification: Confirmed