Columbia / TRI

Diffusion Policy↗

Diffusion Policy is a Baseline from Columbia / TRI, released March 7, 2023. It ranks #31 of 36 overall on the Robot Intelligence Index, with a rating of 1210. Its best result is #4 of 7 on ArmnetBench. Visuomotor policy that generates action sequences by conditional denoising diffusion; standard single-task baseline.

Release date: Mar 07, 2023

Type
Baseline
Model weights
Open
License
MIT
Parameters
—
Pricing
Self-hosted
Benchmarks
4
Overall rank
#31of 36
Rating 1210 · range 20–35
Real robots rank
#38of 56
Rating 1397
Simulation rank
#48of 49
Rating 1006 · range 30–47
Rating
1210± 24.1
10601989
Cost / task · est.
$0.028
$0.0087$17.74
Time / task · est.
31s
9.5 s1 h 22 min

Per task attempt; the tiles show the median across benchmarks. Self-hosted compute is priced at one H100 at $3.29 per hour (Lambda on-demand, Oct 2026).

Inference

~0.1 s inference latency on RTX 3080 with 10 DDIM steps; predicts 16 steps, executes 8 (10 Hz control). OFT paper: 90 ms on A100 with chunk 24 (ALOHA).