Embedl-optimized
- Physical-Intelligence/openpi
- FP8 + FP16
- AGX Thor
- Labels
- Text
- VLA
- Robot Policy
- Build
- Embedl-optimized
- Precision
- FP8 + FP16
- Runtime
- TensorRT 10.16.2.10 with Embedl optimizations
- Action chunk
- 10 actions, 10 denoising steps
π0.5 is Physical Intelligence's vision-language-action policy: it takes camera images and a task instruction and outputs robot actions. This is the official OpenPI LIBERO checkpoint.
Simulator: LIBERO
8f1084e · MuJoCo 3.9.0 · Franka Emika Panda · 20 steps per second
4 scenes
Kitchen counter libero_goal/0
LIBERO-Goal · at most 300 steps
- open the middle drawer of the cabinet
- put the bowl on the stove
- put the wine bottle on top of the cabinet
- open the top drawer and put the bowl inside
- put the bowl on top of the cabinet
- push the plate to the front of the stove
- put the cream cheese in the bowl
- turn on the stove
- put the bowl on the plate
- put the wine bottle on the rack
Bowls and a plate libero_spatial/0
LIBERO-Spatial · at most 220 steps
- pick up the black bowl between the plate and the ramekin and place it on the plate
Groceries and a basket libero_object/0
LIBERO-Object · at most 280 steps
- pick up the alphabet soup and place it in the basket
Living room libero_10/0
LIBERO-10 · at most 520 steps
- put both the alphabet soup and the tomato sauce in the basket
OpenPI code is Apache-2.0 and LIBERO code is MIT; the weights and tokenizer are Gemma Model Derivatives under the Gemma Terms of Use (ai.google.dev/gemma/terms). Not an official Physical Intelligence product.
Methodology
- Latency: one policy call producing a 10-action chunk with 10 denoising steps, batch 1, median of 30 runs on the device.
- Accuracy: closed-loop LIBERO episodes on the page's four scenes, 65 episodes paired with the same OpenPI checkpoint in PyTorch BF16. Init state 0 supplies calibration, state 1 fidelity checks, states 2–4 are the held-out set. A small paired sample, not the full LIBERO benchmark.
- Baseline TensorRT: the same checkpoint in FP16 on upstream TensorRT, same device: not measured yet.
- Precision: the action expert and language MLPs 8–16 use FP8; vision, language attention, MLPs 0–7 and boundary projections stay FP16 to preserve gripper transitions.