π0.5
Physical IntelligenceDuring inference, we use simple greedy autoregressive decoding
general family · filed under inference & serving
Selects the highest-scoring next output without sampling; evidence includes autoregressive generation and action generation.
The sources on this page only mention it in passing. The line above is the extraction's one-line gloss. See decoding strategy for the methods filed beside it.
Also called Greedy action generation without sampling, do_sample=False, Greedy autoregressive decoding, simple greedy autoregressive decoding, Greedy decoding for benchmark generation, greedy decoding without reasoning.
4 quotes from 4 sources, one card per model, each opening with its strongest.
During inference, we use simple greedy autoregressive decoding
Generation: Greedy decoding, max 8192 tokens.
For efficiency reasons, judge requests use greedy decoding without reasoning.
action = vla.predict_action(**inputs, unnorm_key="bridge_orig", do_sample=False)