π

π0.7

Physical Intelligence, United States · April 2026

Foundation VLAClosed weightsAlso written pi 0.7, pi07, pi-zero-point-seven
Parameters
5 B
The paper states about 5B total parameters: a Gemma3 4B VLM backbone (with 400M vision encoder) and an 860M flow matching...
GPU memory
9.3 GB
weights at bf16, computed. Fits a 16 GB card
Inference latency
not published
per action step
Weights
closed
no public checkpoint

What it is

π0.7 is Physical Intelligence's generalist robot foundation model, released on 16 April 2026 with an 87 author technical report. It is roughly 5B parameters: a Gemma3 4B VLM backbone including a 400M vision encoder, plus an 860M flow matching action expert trained with the knowledge insulation recipe. The distinguishing idea is steerability through diverse context conditioning, meaning the model is trained not only on language commands but also on strategy metadata, speed and quality labels, control modality labels and visual subgoal images, which lets an operator shape how a task is done without fine-tuning. The strongest published evidence is zero-shot cross-embodiment shirt folding on a bimanual UR5e that matches human teleoperators. Weights are closed and π0.7 is not in the openpi repository, which still ships only π0, π0-FAST and π0.5.

Architecture

Backbone
Gemma3 4B VLM, including a 400M parameter vision encoder, plus a MEM-style video history encoder with temporal and spatial compression
Action head
Flow matching action expert (860M parameters), trained with the knowledge insulation recipe so action-expert gradients do not flow into the VLM, 5 denoising steps at inference
Action chunk
50 steps per forward pass
Parameters
The paper states about 5B total parameters: a Gemma3 4B VLM backbone (with 400M vision encoder) and an 860M flow matching action expert. Verbatim: the model consists of a VLM backbone initialized from the Gemma3 4B-parameter VLM, and a flow matching action expert with 860M parameters.
Pretraining data
Not quantified in hours. The paper lists teleoperated demonstrations from static, mobile, single-arm and bimanual platforms in labs, homes and in the wild, plus autonomous data from policy evaluations including failures, human intervention data inside policy trajectories, open source robot datasets, egocentric human video, and auxiliary web data for object localization, VQA, text prediction and video captioning.
Embodiments
bimanual mobile manipulators with dual 6-DOF arms, static bimanual BiPi platform with lightweight 6-DOF arms, bimanual UR5e with Robotiq grippers, single-arm 6-DOF systems, Franka via DROID

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 5 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
fp3219 GBTraining master weights, rarely used for inference
bf16 or fp169.3 GBThe usual way these checkpoints are served
int84.7 GBQuantised, expect some loss of precision on fine motions
int42.3 GBAggressive quantisation, verify success rate before trusting it

Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 16 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.

Not published. The paper states the robot control rates rather than a model property: the UR5e platforms run at 20 Hz and all other robots at 50 Hz, and the results listed here were collected on the UR5e.

The authors publish no memory or latency figure for this model. Everything above is computed from the parameter count.

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

No results in this category are published for this model.

Fine tuned tasks

What the model reaches after being adapted to a new task or a new robot, usually from a small number of demonstrations. This is the number that matters if you bring your own data.

  • Physical Intelligence real-robot evaluation Zero-shot cross-embodiment shirt folding on bimanual UR5e, no UR5e laundry data in training
    85.6%
    task progressAuthors' own evaluation. Human expert teleoperators reached 90.9 percent task progress.source
  • Physical Intelligence real-robot evaluation Zero-shot cross-embodiment shirt folding on bimanual UR5e, no UR5e laundry data in training
    80%
    success rateAuthors' own evaluation. Human expert teleoperators on the same unfamiliar hardware reached 80.6 percent.source

On real hardware

All evaluations are on real hardware. Most robots run at 50 Hz, the UR5e platforms at 20 Hz. Zero-shot cross-embodiment shirt folding on a bimanual UR5e with no UR5e laundry data in training reaches 85.6 percent task progress and 80 percent success, against 90.9 percent progress and 80.6 percent success for human expert teleoperators on the same setup. On espresso making, box building and laundry folding the model matches task-specific RL-trained specialist policies and exceeds them in throughput on laundry and box building. Instruction following was tested in 14 scenarios across 4 unseen kitchens and 2 unseen bedrooms.

Fine tuning it yourself

Not applicable publicly. The central claim is that π0.7 reaches specialist-level performance out of the box without task-specific fine-tuning, and no weights or fine-tuning code are released. openpi covers π0, π0-FAST and π0.5 only.

Where it helps, where it does not

Strengths

  • Steerability through multimodal prompting: strategy metadata, speed and quality labels, control modality labels and visual subgoal images let an operator change how a task is executed without retraining.
  • Documented architecture with an exact parameter split, which most closed competitors do not provide.
  • Zero-shot cross-embodiment transfer that reaches human teleoperator level on an unfamiliar bimanual UR5e.
  • Matches task-specific RL specialists on espresso making, box building and laundry folding without task-specific fine-tuning, and beats them on throughput for two of the three.
  • Trained to tolerate asynchronous inference delay of up to 12 timesteps, which is a real deployment concern at 50 Hz.

Limits

  • Weights and code are closed. openpi ships π0, π0-FAST and π0.5, not π0.7.
  • The authors state seen tasks often exceed 90 percent success while unseen tasks or unseen task and robot combinations land in the 60 to 80 percent range.
  • The authors concede it is practically difficult on a dataset this large to determine which tasks are truly seen or unseen, which weakens the zero-shot framing.
  • No measured inference latency, VRAM figure or minimum GPU is published. The 240 ms number in the paper is a simulated training-time delay, not a measurement.
  • Training data volume is never quantified in hours, so the data scale cannot be compared against LingBot-VLA 2.0 or GEN-1.
  • All reported embodiments use parallel jaw grippers. No multi-finger hand results are reported.
  • The blog notes some zero-shot tasks still need language coaching or a fine-tuned high-level policy to reach full competency.

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.