GR-3

ByteDance Seed · 2025-07

VLAIndustry models4Bflow matching

Co-trains on robot trajectories, web vision-language data and VR human trajectories to generalize to new objects·instructions.

Paper

Backbone LLM: Qwen2.5-3B · Qwen2.5-VL-3B config
Vision encoder: Qwen2.5-VL ViT
Action head: flow matching action DiT (MoT) (flow matching)
Chunk: k steps (not disclosed)
Control rate: Not disclosed
Architecture (original summary): Qwen2.5-VL-3B-Instruct + action DiT (half the VLM's layers), mixture-of-transformers; flow matching, k-length action chunk represented as k tokens (k not disclosed), causal attention in action DiT

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 + web vision-language + motion capture (VR human trajectories)
Data scale: pick-and-place 35k traj·101 objects·69 h; table tidying about 101 h; cloth handling 116 h; VR human trajectories about 450 per hour (teleoperation about 250), up to 10 per object for 45 unseen objects
Main datasets: In-house ByteMini teleoperation data, web VL (captioning, VQA, grounding, interleaved grounded captioning), PICO 4 Ultra Enterprise VR human trajectories
Tasks & evaluation: pick-and-place unseen instructions 77.1%, unseen objects 57.8% (86.7% with 10-shot human trajectories); table tidying instruction following 97.5% vs π0 53.8%; cloth handling average task progress 86.7/83.9/75.8%
Robot: ByteMini 22-DoF bimanual mobile robot (7-DoF arms, spherical wrists, RGBD on head · both wrists)

Inference device & speed: Uses only the KV cache of the latter half of VLM layers; hardware·latency·control frequency not disclosed
Inference memory: Not disclosed
Weights bf16 (calc.): 8 GB
Where it runs: Not disclosed

Reinforcement learning: Not used · Imitation learning + co-training with web VL·VR human trajectories. The RL-applied version is GR-RL
Base control: Wheeled (Direct base velocity/pose output) · Policy chunks control the robot end-to-end including the mobile base (19 DoF excluding lift and head); base moves and rotates in table cleanup and clothes hanging

Key features: Co-trains on robot trajectories, web vision-language data and VR human trajectories to generalize to new objects·instructions. Adds RMSNorm after attention·FFN linear layers to stabilize training.
Notes: Data hours are stated only per task (69 h, about 101 h, 116 h) with no total, so the numeric cell is left empty. Organization on arXiv page 1 is 'ByteDance'. No public repository found (could not be confirmed in verification). arXiv v1 2025-07-21.
Verification: Confirmed