AY-ROBOTSAI Teleoperation
Sign Up
π

π*0.6

Physical Intelligence, United States · November 2025

Reasoning VLAClosed weightsAlso written pistar06, pi*0.6, pi-star-0.6, RECAP
Parameters
not published
No total is published.
GPU memory
not published
weights at bf16, computed
Inference latency
not published
per action step
Weights
closed
closed.

What it is

π*0.6 is not a new architecture so much as a training method applied to one. The underlying π0.6 model upgrades π0.5 with a Gemma 3 4B backbone and an 860M action expert, and π*0.6 adds conditioning on a binarised advantage signal produced by a separate 670M value model. RECAP then improves the policy from three data sources at once: demonstrations, autonomous rollouts, and teleoperated corrections during those rollouts. The results are the strongest published evidence that VLAs can learn from deployment, roughly doubling throughput and halving failures on the hardest tasks. For anyone making a build decision the caveats dominate: nothing is released, the headline numbers exist only as bar charts, and there is no public benchmark result to compare against.

Architecture

Backbone
Gemma 3 4B, the base VLM of the underlying π0.6 model. A second, smaller 670M Gemma 3 initialised VLM serves as the value function backbone.
Action head
flow matching action expert (860M) trained under the Knowledge Insulation recipe with a stop gradient, alongside autoregressive prediction of a next-subtask text token sequence and of FAST-tokenised actions. At inference the policy is additionally conditioned on a binarised advantage indicator.
Control rate
50 Hz
Parameters
No total is published. The paper gives the components: the base VLM is Gemma 3 4B and the action expert was increased to 860M parameters, up from 300M in π0 and π0.5. The separate distributional value function uses its own 670M VLM backbone with B=201 value bins. The blog refers loosely to a 5B parameter vision-language model. No total for the deployed policy is stated anywhere, so it is left null rather than summed.
Pretraining data
Follows the π0.5 recipe: web vision-language data, subtask prediction and low-level actions from many tasks and many robots, but pre-trained with offline RL rather than plain imitation. The value function is pre-trained on the same dataset to predict the negative number of steps to task completion, and the per-task improvement threshold is set to the 30th percentile of predicted values. Downstream, RECAP adds task-specific demonstrations plus one or more rounds of on-robot autonomous rollouts and expert teleoperated interventions.
Embodiments
bimanual mobile manipulator platform used for laundry folding in real homes, bimanual station used for espresso making with a professional espresso machine, bimanual station used for cardboard box assembly in a packaging setting

What hardware it needs

This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.

Published figures. These come from the model authors, not from this site.
Control rate50 Hz

Not published. The model is closed, so no hardware requirements are documented.

Reported results

Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.

Simulation

No results in this category are published for this model.

Real world

Rollouts on physical hardware. The setups differ, so read these as evidence, not as a ranking.

  • Physical Intelligence real-robot RECAP evaluation Laundry folding after the first on-robot RL iteration
    90%
    success rateFirst-party. The paper says the first iteration already raises the success rate to over 90 percent, so 90 is a floor and the exact value is only shown in a chart. The second iteration improves throughput rather than success rate.source
  • Physical Intelligence real-robot RECAP evaluation Box assembly, folding and labelling within a 600 second time limit, final policy
    90%
    success rateFirst-party. Paper states about 90 percent for both folding the box and labelling it within the allotted 600 seconds. Placing the finished box on the stack remains a residual failure mode.source
  • Physical Intelligence real-robot RECAP evaluation Espresso making, uninterrupted autonomous operation
    13 hours
    continuous run time without interventionFirst-party robustness demonstration, not a comparative benchmark. The blog describes the same run as serving drinks from 5:30am to 11:30pm. Laundry folding ran over two hours uninterrupted on 50 novel items in a new home.source
  • Physical Intelligence real-robot RECAP evaluation Diverse laundry folding, final π*0.6 versus the offline RL plus SFT model
    2 x improvement
    throughput, successfully completed tasks per hourFirst-party. Paper text states throughput more than doubles on diverse laundry folding and espresso. The absolute per-hour figures appear only in bar charts (Figure 7) and are not tabulated.source
  • Physical Intelligence real-robot RECAP evaluation Espresso making, final π*0.6 versus the offline RL plus SFT model
    2 x improvement
    throughput, successfully completed tasks per hourFirst-party. Reported as more than doubling. Failure rate falls by roughly a factor of two on the same task. Absolute values only in bar charts.source

Fine tuned tasks

No results in this category are published for this model.

On real hardware

Three long-horizon real tasks, each 5 to 15 minutes per episode: making espresso drinks on a professional espresso machine, folding diverse laundry in real homes, and assembling cardboard boxes used for real packaging. Reported endurance runs cover 13 hours of continuous espresso service, more than two hours of uninterrupted laundry folding on 50 novel items in a new home, and 59 boxes assembled for a chocolate packaging factory. Success is judged by human raters aggregating several quality indicators.

Fine tuning it yourself

Not applicable to outside users, since nothing is released. The paper's own procedure is a three-stage pipeline: offline RL pre-training of π*0.6 on the multi-task multi-robot corpus, supervised fine-tuning on task demonstrations, then one or more RECAP iterations of autonomous on-robot rollouts with expert interventions, retraining the value function and the advantage-conditioned policy each round. Two iterations gave a 50 percent throughput gain on laundry and a 2x gain on box assembly.

Where it helps, where it does not

Strengths

  • The first published demonstration that a general RL recipe with human reward feedback and teleoperated interventions measurably improves a large VLA from its own deployment experience, rather than from more demonstrations.
  • Advantage conditioning sidesteps the tractable-log-likelihood problem that blocks PPO and REINFORCE on flow matching policies, and it trains on all data instead of discarding low-advantage samples the way AWR does.
  • Beat both AWR and PPO baselines on the same collected data in a controlled comparison, with PPO needing a tight trust region (η = 0.01) just to stay stable.
  • Optimises throughput, meaning successful completions per hour, not only success rate. That is the metric that decides whether a robot is economically useful.
  • Robustness is shown in hours of continuous unattended operation, not in episode counts.

Limits

  • Completely closed. No weights, no code, no API, no checkpoint size and no hardware requirements. Nothing here can be reproduced or deployed by a third party.
  • All headline numbers appear as bar charts with standard error, not as tables, so the exact throughput and success values cannot be quoted precisely.
  • No total parameter count is published, and no inference latency or VRAM figure exists at all.
  • Evaluated on three tasks on Physical Intelligence's own hardware. No public simulation benchmark, no LIBERO, no SimplerEnv, so there is no comparable number against any open model.
  • The method needs on-robot data collection with human supervisors providing interventions and reward annotation. That operational cost is the real barrier, and the paper does not quantify how many intervention hours each task took.
  • Already superseded. π0.7 (April 2026) is described by the vendor as matching these RL-trained specialists with a single model and no task-specific fine-tuning.

Sources

Everything on this page was taken from these documents. Where they disagree with what you read here, they win.

Entry last checked 2026-08-11.

Try a policy on a real arm

A physical SO-100 is online and free to drive in the browser, and five of the models in this arena can be fine tuned on a dataset you record with your own.

Download the clientDrive a real arm now