VLA Gallery / All models / Reference baselines (pre-2024)

Diffusion Policy (CNN)

Columbia University; Toyota Research Institute; MIT; Stanford University · 2023-03

VAReference baselines (pre-2024)Denoiser 256Mdiffusion

Policy that generates action sequences by diffusion denoising to handle multimodal action distributions.

Paper

Category: Reference baselines (pre-2024)
Type: VA
Released: 2023-03
Organization: Columbia University; Toyota Research Institute; MIT; Stanford University
Open weights: Yes
Parameters: Denoiser 256M (sim, Transport 264M)·67M (real tasks) + vision encoder 22M (Transport 45M, state-based 0)

Backbone LLM: None
Vision encoder: ResNet-18
Action head: 1D CNN U-Net diffusion (DDIM 10–16 steps) (diffusion)
Chunk: Predict 16 · execute 8
Control rate: 10 Hz
Architecture (original summary): ResNet-18 (no pretraining, spatial softmax, GroupNorm) + FiLM-conditioned 1D temporal CNN U-Net; DDPM action sequence generation (To 2, Ta 8, Tp 16), 100 training steps·DDIM 10–16 inference steps, real 10 Hz

Pretraining accelerators: N/A
Pretraining duration: N/A
Pretraining steps & batch: N/A
GPU/TPU-hours: Not disclosed
Fine-tuning: Per task: training GPU Not disclosed (paper·README); 3000 epochs image·4500 state, batch 64 (image)·256 (state), lr 1e-4, warmup 500; real Push-T fixed 12 hours per method; (secondary) re-measured in CARP paper: V100 12.17 GPU-hours (state-based Square, 200 trajectories)
Data type: Simulation·synthetic; real-robot teleoperation
Data scale: Robomimic 200 (proficient)·300 (mixed) demos per task, Push-T 200, Block Push 1,000 (scripted), Kitchen 566; real Push-T 136, source task 50 (90% train), Egg Beater 210, Mat Unrolling 162, Shirt Folding 284; Mug Flipping not disclosed†
Main datasets: Robomimic; Push-T; Block Push; Franka Kitchen; in-house real demos
Tasks & evaluation: Robomimic Lift·Can·Square·Transport·ToolHang, Push-T, Block Push, Franka Kitchen; real Push-T (UR5), Mug Flipping, Sauce Pouring·Spreading (Franka), bimanual Egg Beater·Mat Unrolling·Shirt Folding (Franka)
Robot: UR5 (real Push-T); Franka (2 source tasks, 3 bimanual tasks); Mug Flipping robot not specified in notes

Reinforcement learning: Not used · Imitation learning (diffusion behavior cloning)
Base control: Fixed · No base control (UR5·Franka arms)

Key features: Policy that generates action sequences by diffusion denoising to handle multimodal action distributions. Many 2024–2025 VA papers use it as a standard comparison baseline.
Notes: Parameters (B) is the sum of #D 264M + #V 45M for the largest setting, Transport (listed separately in the paper). All real-task rows are #D 67M, #V 22M. Sim checkpoints released in the training log folder. Journal version v5 (2024-03) added bimanual tasks. Stanford added to affiliations (correction applied).
Verification: Corrections applied