Data Scaling Laws

Tsinghua University; Shanghai Qi Zhi Institute; Shanghai AI Lab · 2024-10

VAVAOver 396Mdiffusion

Collected over 40k demonstrations with UMI and showed that generalization follows a power law in the number of environments·objects.

Paper

Category: VA
Type: VA
Released: 2024-10
Organization: Tsinghua University; Shanghai Qi Zhi Institute; Shanghai AI Lab
Open weights: Yes
Parameters: Over 396M (vision encoder + noise prediction network)

Backbone LLM: None
Vision encoder: DINOv2 ViT-L/14 (fine-tuned)
Action head: CNN U-Net diffusion (diffusion)
Chunk: 16 steps
Control rate: 5 Hz
Architecture (original summary): Fine-tuned DINOv2 ViT-L/14 + CNN U-Net diffusion policy; action horizon 16, image observation horizon 2 (Pour Water·Unplug Charger 3), environment frequency 5

Pretraining accelerators: N/A
Pretraining duration: N/A
Pretraining steps & batch: N/A
GPU/TPU-hours: Not disclosed
Fine-tuning: Per task: largest-dataset policy 8× A800, 75 epochs (5×10^5 steps), 75 hours; batch 256
Data type: Human video (UMI handheld gripper demonstrations)
Data scale: 40,000+ demonstrations total; released policy uses 32 environment-object pairs × 50 demonstrations per environment
Main datasets: In-house collection with UMI
Tasks & evaluation: Pour Water, Mouse Arrangement, Fold Towels, Unplug Charger; 15,000+ real-robot rollouts; about 90% success on 2 tasks in new environments·with new objects using half a day of data from 4 collectors
Robot: Franka Emika Panda + Weiss WSG-50, wrist GoPro

Reinforcement learning: Not used · Imitation learning from human handheld demonstrations
Base control: Fixed · No base control (Franka arm)

Key features: Collected over 40k demonstrations with UMI and showed that generalization follows a power law in the number of environments·objects. Concluded that diversity matters more than data quantity.
Notes: Parameters (B) is a lower bound (over 396M). Single-task policies, no language. Four task policies released on Hugging Face. ICLR 2025.
Verification: Confirmed