π

π0.5

Physical Intelligence, United States · April 2025

Foundation VLAOpen weights, Apache-2.0Also written pi0.5, pi05, pi oh fiveTrainable on AY-Robots
Parameters
3.6 B
The π0.5 paper does not state a parameter count anywhere.
GPU memory
8 GB
reported by the authors
Inference latency
not published
per action step
Weights
Apache-2.0
Apache-2.0 for the openpi repository.

What it is

π0.5 keeps π0's architecture and changes what it is fed. Pre-training runs on FAST discrete tokens across mobile robots, static robots, cross-embodiment lab data, subtask labels, verbal coaching and plain web data, and only the post-training stage bolts on the flow matching expert for fast continuous chunks. The striking number is that 97.6 percent of first-stage examples are not mobile manipulation, and the model still cleans kitchens in homes it has never seen. It also carries the one directly comparable public score in the family, 96.85 percent average on LIBERO from the openpi repository. The gap in the record is specification detail: the paper states no parameter count, no backbone name and no latency, so those have to be read out of the code.

Architecture

Backbone
PaliGemma-3B, same backbone family as π0. The paper does not name the VLM, but openpi implements π0.5 through the same Pi0Config with the pi05 flag set, using paligemma_variant gemma_2b and action_expert_variant gemma_300m.
Action head
two stage. Pre-training represents every action as FAST discrete tokens for scalable next-token training. Post-training adds a flow matching action expert initialised from scratch for continuous chunks at inference. The model also autoregressively emits a high-level subtask in text, then conditions the action generation on it.
Action chunk
50 steps per forward pass
Control rate
50 Hz
Parameters
The π0.5 paper does not state a parameter count anywhere. Two independent primary sources bracket it. First, openpi implements π0.5 with the identical Pi0Config used for π0 (paligemma_variant gemma_2b plus action_expert_variant gemma_300m), which the π0 paper counts as 3.3 billion total. Second, the LeRobot mirror lerobot/pi05_base reports 3,616,757,520 parameters in its safetensors metadata. The blog names only the action expert, at 300M.
Pretraining data
Deliberately heterogeneous co-training mixture. About 400 hours of mobile manipulator data collected in about 100 real home environments (104 locations in the final model), plus non-mobile single-arm and bimanual robot data from many homes (ME), cross-embodiment lab data (CE), high-level subtask prediction data (HL), verbal instructions from human supervisors (VI), and web data covering captioning, visual question answering and object localisation (WD). 97.6 percent of the first-stage training examples do not come from mobile manipulators doing household tasks. Pre-training runs 280k gradient steps before post-training.
Embodiments
mobile manipulators with two arms, a torso lift and a mobile base, static single-arm robots, static bimanual robots, Franka (via the released π0.5-DROID checkpoint)

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.

Computed from 3.6 B parameters. Weights only, so treat it as a floor: activations, image encoder intermediates and the CUDA context come on top, in practice 30 to 50 percent more for inference.
PrecisionWeightsNote
fp3213 GBTraining master weights, rarely used for inference
bf16 or fp166.7 GBThe usual way these checkpoints are served
int83.4 GBQuantised, expect some loss of precision on fine motions
int41.7 GBAggressive quantisation, verify success rate before trusting it

Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 12 GB. That is an estimate for inference, not for fine tuning. Full fine tuning also holds optimiser state and gradients, which typically costs several times the weights unless you use LoRA or a comparable adapter.

Published figures. These come from the model authors, not from this site.
Inference memory, reported8 GB
Fine tuning memory, reported22.5 GB
Control rate50 Hz

openpi names the RTX 4090 as the example GPU for inference and LoRA fine-tuning, and an A100 80GB or H100 for full fine-tuning.

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

Reproducible benchmark suites. Everyone runs the same episodes, so the numbers can be compared inside a suite.

  • LIBERO Spatial, fine-tuned checkpoint at 30k steps
    98.8%
    success rateFirst-party.source
  • LIBERO Object, fine-tuned checkpoint at 30k steps
    98.2%
    success rateFirst-party.source
  • LIBERO Goal, fine-tuned checkpoint at 30k steps
    98%
    success rateFirst-party.source
  • LIBERO Average over Spatial, Object, Goal and Long, fine-tuned checkpoint at 30k steps
    96.85%
    success rateFirst-party, reproducible from the pi05_libero checkpoint trained with the openpi pi05_libero config.source
  • LIBERO Long (LIBERO-10), fine-tuned checkpoint at 30k steps
    92.4%
    success rateFirst-party. The long-horizon suite is the weakest of the four, which is the usual pattern.source

Real world

No results in this category are published for this model.

Fine tuned tasks

No results in this category are published for this model.

On real hardware

Evaluated in three real kitchens and three real bedrooms in homes not present in the training set, using both a mobile manipulator and a static robot, on tasks including dishes in sink, items in drawer, laundry basket and make bed. Long-horizon runs of 10 to 15 minutes clean an entire kitchen or bedroom from one high-level prompt. The control stack is deliberately plain: the model directly commands arm, gripper and torso-lift targets plus base velocities at 50 Hz with action chunking, tracked by simple PD controllers, with no trajectory planning and no collision checking.

Fine tuning it yourself

Fine-tune through openpi. Convert data to the LeRobot dataset format, then define LiberoInputs and LiberoOutputs style transforms, a LeRobotDataConfig and a TrainConfig. The repository ships the pi05_libero config as a complete worked example, including compute_norm_stats and a Dockerised evaluation loop. π0.5-DROID is offered as a fine-tuning base for Franka setups and uses the knowledge insulation recipe. No minimum demonstration count is documented.

This is one of the models the AY-Robots training pool can fine tune for you on a dataset recorded with your own arm. The platform page carries the defaults, GPU tier and inference latency measured in that pool.

Platform page for π0.5

Where it helps, where it does not

Strengths

  • The only Physical Intelligence model with open weights that also has a published public-benchmark number, at 96.85 percent average on LIBERO, which makes it the one π model that can be compared directly against GR00T and others.
  • Demonstrated end to end in three real homes never seen in training, doing 10 to 15 minute tasks such as clearing a kitchen or a bedroom from a single high-level prompt.
  • The co-training ablation is the useful engineering result. Performance with 104 training locations matches a control model that was trained on data from the test homes themselves, so the recipe closes the generalisation gap rather than memorising environments.
  • Keeps π0's cheap 8 GB inference and 50 Hz chunked control while adding the FAST-token pre-training stage that made training scale.
  • Emits an explicit high-level subtask in text before acting, which gives a debuggable intermediate signal instead of an opaque policy.

Limits

  • The paper never states a parameter count, a backbone name or an inference latency. Every such figure has to be taken from the openpi implementation or a mirror, not from the publication.
  • Real-home results are reported as task progress rubrics averaged over 10 trials per task and environment, not as clean success rates, so they cannot be compared against success-rate numbers from other models.
  • No latency measurement is published for π0.5 specifically. The 73 ms figure belongs to π0 and should not be carried over silently, even though the architectures match.
  • Full fine-tuning still needs more than 70 GB, so a single consumer GPU covers inference and LoRA only.
  • Already two generations behind the vendor's own frontier. π0.6, π*0.6 and π0.7 all followed, and none of them has open weights.

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.