Physical Intelligence

π0 (per-task)↗

π0 (per-task) is a VLA from Physical Intelligence, released October 31, 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 #10 of 22 on RoboChallenge. 3.3B flow-matching VLA: PaliGemma VLM plus a 300M action expert outputting 50-step action chunks. (Fine-tuned separately per task.)

Release date: Oct 31, 2024

Type
VLA
Model weights
Open
License
Apache-2.0
Parameters
3.3B
Pricing
Self-hosted
Benchmarks
1
Overall rank
Not ranked
Needs a simulation result to rank overall.
Real robots rank
#28of 56
Rating 1495
Simulation rank
Not ranked
No simulation results yet.
Rating
1474
10601989
Cost / task · est.
$0.267
$0.0087$17.74
Time / task
4min52s
9.5 s1 h 22 min
Rank on each benchmark
BenchmarkScoreRank
RoboChallengeProgress46.410 / 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 Physical Intelligence

Inference

73 ms per 50-step chunk on-board on RTX 4090 with 3 cameras (14 ms image encoders + 32 ms prefix + 27 ms for 10 flow steps); 86 ms off-board. Robots run at 20-50 Hz, re-planning every 16-25 actions. OFT paper measured 86 ms on A100 (ALOHA).