Physical Intelligence Has Gone Quiet Since π0.7. Here's Who Is Publishing Instead

Physical Intelligence's own blog lists no research post for 180 days after "π0.7: a Steerable Model with Emergent Capabilities" (16 April 2026). The only later entry, dated 13 October 2026, is a product announcement for an API beta, not a paper. Other groups kept publishing robot foundation models in the gap, including A1, Xiaomi-Robotics-1 and PhysBrain 1.5.

Has Physical Intelligence really gone quiet?

On its blog, yes, for research. As of 14 October 2026 the blog listing shows seven posts between 17 November 2025 and 16 April 2026, then nothing until 13 October 2026.

Type labels are this site's reading of the titles. The caveat matters: a quiet blog is not proof of a quiet lab. No source gives a reason for the gap, and the listing does not show papers or demos published elsewhere.

What was the last thing Physical Intelligence published?

It was π0.7, posted 16 April 2026, with a technical report on arXiv as 2604.15483 (v2 revised 24 April 2026). The arXiv page credits "Physical Intelligence and 87 other authors."

According to the authors, π0.7 is a 5B-parameter model with a Gemma3 4B backbone, a video history encoder and an 860M-parameter flow-matching action expert. Its core idea is diverse context conditioning. The prompt carries not only a language command but also subgoal images and episode metadata (speed, a 1 to 5 quality score, a mistake flag), so that failures and mediocre demonstrations can be used as labelled training data instead of being discarded.

The authors report that the model folds laundry on a bimanual UR5e without having seen the task on that robot. In their human comparison, The Decoder reports an 80 percent success rate for the model, which the paper's HTML puts against 80.6% for expert teleoperators over 30 trials. These are lab-reported numbers.

What did Physical Intelligence announce on 13 October 2026?

It announced a Pi API beta. The company says users can "upload data, launch fine-tuning on compute and infrastructure we manage, and access fine-tuned models via a hosted gateway to run inference."

Details the post does and does not give:

• Access is through a waitlist. No pricing is published.
• It says "the latest model" is available but does not name a version.
• Partner examples include Actor (an autonomous wheel loader in Utah) and Almond, whose dual-arm robot Axol ran a model obtained after four days of iterating with the API. These are company-reported demos.
• The post lists failure cases (an inside-out garment, an ingredient that would not stay in a scoop) and admits network connectivity problems for some early deployments.

The post contains no benchmarks, methods or ablations. It reads as a shift in what the company talks about publicly: from papers toward a service. That is the site's interpretation, not something the company states.

Who is publishing robot foundation models instead?

Several groups posted large VLA papers in the same window. Here are three, with each paper's own headline claim.

A1 is an efficiency paper. Its abstract claims up to 72% lower per-episode latency for flow-matching inference and up to 76.6% less backbone computation, with "minor performance degradation."

Xiaomi-Robotics-1 is a data-scale paper. It pre-trains on over 100,000 hours of trajectories collected with UMI handheld devices, with language labels added by an automatic pipeline. Its three variants range from 2.6B to 10.5B total parameters. The authors report 75% average success on four dexterous real-robot tasks after fine-tuning, against 40% for π0.5. Its scaling experiments use only about 20,000 of the 100,000 hours.

PhysBrain 1.5 takes a different route. It trains one 8B model, started from Qwen3-VL, to understand scenes, emit end-effector actions and predict future frames, all as next-token prediction. Embodied pre-training supervision comes entirely from human interaction videos. The action and future-frame outputs are shown as qualitative examples, so its 72.5 average measures understanding, not robot control.

Are these papers comparable to π0.7?

Not directly. They use different benchmarks, different robots and different baselines, and none of the three reports a head-to-head against π0.7 in their abstracts.

• A1 and Xiaomi-Robotics-1 compare against π0 and π0.5, not π0.7.
• PhysBrain 1.5 is evaluated mainly on understanding benchmarks, so it answers a different question.
• Each paper picks the baselines that appear in its table. A new state of the art on RoboCasa365 or RoboChallenge says little about how a system behaves on a different robot.

One trend is visible, though. The open or soon-to-be-open releases lean on scale of data (Xiaomi), efficiency (A1) or human video (PhysBrain). π0.7's own pitch was steering through richer prompts and metadata. That is a difference in recipe, and the available evidence does not rank the recipes.

What does a working researcher take from this?

Three practical points.

1. Do not treat blog silence as a signal about method quality. The blog shows what a company chooses to publish, not what it is building.
2. If you need open code today, A1 is the one with a stated full release. Xiaomi-Robotics-1 has promised code and checkpoints, which is not the same as having them. Check the repositories before planning around either.
3. Read the baselines column first. A 0.67-point margin over pi0 on RoboChallenge (29.00% vs 28.33%, from A1's abstract) is small, and the abstract gives no trial counts or confidence intervals.

For the π0.7 design itself, the metadata-conditioning idea is the transferable part: label data by quality and speed, then condition on those labels instead of filtering.

What's still unclear?

• Why the blog gap exists. No source states a reason. It may reflect a product focus, work not yet released, or nothing at all.
• Whether the API serves π0.7. The post says only "the latest model."
• How novel π0.7's tasks were. The authors concede it is hard to determine with certainty which tasks are truly novel given the dataset's size and diversity. The Decoder notes that the air-fryer demo sits close to training data (two episodes of a robot closing an air fryer, plus DROID data with a Franka arm), and PI calls those episodes "quite different." Separating real generalization from recombination is, in The Decoder's framing, the central evaluation problem.
• Reading depth. The π0.7, Xiaomi and PhysBrain claims here rest on the abstracts plus parts of each paper's HTML text; A1's rest on its abstract only. Full PDFs, complete results tables, limitations sections and trial counts are not covered.
• Replication. None of the numbers above has been independently reproduced.