Microsoft Research Asia

CogACT (per-task)↗

CogACT (per-task) is a VLA from Microsoft Research Asia, released November 29, 2024. It has no overall rank on the Robot Intelligence Index, which needs both a simulation and a real-robot result. Its best result is #19 of 22 on RoboChallenge. VLA with a 7B VLM producing a cognition feature that conditions a diffusion-transformer action module. (Fine-tuned separately per task.)

Release date: Nov 29, 2024

Type
VLA
Model weights
Open
License
MIT
Parameters
—
Pricing
Self-hosted
Benchmarks
1
Overall rank
Not ranked
Needs a simulation result to rank overall.
Real robots rank
#47of 56
Rating 1287
Simulation rank
Not ranked
No simulation results yet.
Rating
1264
10601989
Cost / task · est.
$0.202
$0.0087$17.74
Time / task
3min41s
9.5 s1 h 22 min
Rank on each benchmark
BenchmarkScoreRank
RoboChallengeProgress21.819 / 22

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).

Other models from Microsoft Research Asia

Inference

Predicts current + 15 future actions with a diffusion transformer head. EfficientVLA reports ~134.5 ms language module + ~51.5 ms action module (10 denoising steps); GPU not stated.