Physical Intelligence's π0.7 robot model performs tasks outside its training data
Physical Intelligence published research showing its π0.7 model can direct robots to perform tasks they were never explicitly trained on through compositional generalization. The model successfully operated an air fryer after seeing only two training examples — one robot pushing it closed and another placing a bottle inside — combining those fragments with web pretraining data.
Physical Intelligence's π0.7 robot model performs tasks outside its training data
Physical Intelligence published research Thursday showing its π0.7 model can direct robots to perform tasks they were never explicitly trained on, according to the San Francisco-based robotics startup. The capability, which the company's researchers say surprised them, represents what they describe as compositional generalization — combining skills learned in different contexts to solve new problems.
The model successfully operated an air fryer with only two relevant training examples: one where a different robot pushed the appliance closed, and one from an open-source dataset where another robot placed a plastic bottle inside. With zero coaching, π0.7 made what researchers called "a passable attempt" at cooking a sweet potato. With step-by-step verbal instructions, it performed successfully.
"Once it crosses that threshold where it goes from only doing exactly the stuff that you collect the data for to actually remixing things in new ways, the capabilities are going up more than linearly with the amount of data," says Sergey Levine, co-founder and UC Berkeley professor. "That much more favorable scaling property is something we've seen in other domains, like language and vision."
Performance and limitations
Physical Intelligence measured π0.7 against its own previous specialist models — purpose-built systems trained on individual tasks — and claims the generalist model matched their performance across tasks including making coffee, folding laundry, and assembling boxes. The company notes standardized benchmarks for robotics don't exist, making external validation difficult.
The model cannot yet execute complex multi-step tasks autonomously from a single high-level command. "You can't tell it, 'Hey, go make me some toast'," Levine says. "But if you walk it through — 'for the toaster, open this part, push that button, do this' — then it actually tends to work pretty well."
Prompt engineering significantly affected results. Research scientist Ashwin Balakrishna, a Stanford computer science PhD student, says an early air fryer experiment produced a 5% success rate. After refining how the task was explained to the model for about 30 minutes, the success rate jumped to 95%, according to the company.
Research context
The paper uses careful hedging language throughout, describing π0.7 as showing "early signs" of generalization and "initial demonstrations" of new capabilities. When asked about deployment timelines, Levine declined to speculate: "I think there's good reason to be optimistic, and certainly it's progressing faster than I expected a couple of years ago. But it's very hard for me to answer that question."
Physical Intelligence has raised over $1 billion to date at a $5.6 billion valuation. The company is now in discussions for a new funding round that would value it at $11 billion, according to the report. The company declined to comment on fundraising.
What this means
If validated externally, π0.7's compositional generalization would mark a departure from robotics' standard approach of training specialist models on specific tasks through data collection. The ability to coach robots through unfamiliar tasks with verbal instructions could enable deployment in new environments without additional data collection or retraining. However, the lack of standardized robotics benchmarks and reliance on the company's own internal measurements makes independent verification of these claims difficult. The model's heavy dependence on prompt engineering quality and inability to handle complex multi-step tasks autonomously indicate the technology remains in early research stages.
Related Articles
AI2 Releases DiScoFormer: Single Transformer Estimates Density and Score Across Distributions Without Retraining
Allen Institute for AI (AI2) has released DiScoFormer, a transformer model that estimates both the density and score of any distribution from a sample in a single forward pass without retraining. In 100 dimensions, the model reduces score estimation error by 6.5x and density error by 37x compared to classical kernel density estimation.
AI2 Research: Hybrid Models Excel at Content Words, Transformers Better at Token Repetition
Allen Institute for AI researchers conducted token-level analysis comparing their 7B-parameter Olmo 3 transformer and Olmo Hybrid models. The study finds hybrid architectures show a loss gap advantage of 0.04 on content words (nouns, verbs, adjectives) versus 0.02 on function words, while transformers match or exceed hybrids on repeated tokens and closing braces.
Mistral AI fine-tunes Pixtral-12B on satellite imagery, boosting classification accuracy from 56% to 91%
Mistral AI has published research showing that fine-tuning its Pixtral-12B vision language model on satellite imagery increases classification accuracy from 56% to 91% on the Aerial Image Dataset. Using Low-Rank Adaptation (LoRA) with 8,000 training samples across 30 scene categories, the company reduced hallucinations from 5% to 0.1% for under $10 in compute costs.
NVIDIA Shows Task-Seeded Synthetic Data Boosts Nemotron-3 Nano by +11.1 on GPQA
NVIDIA demonstrated that task-seeded synthetic Q&A data improves model performance across multiple benchmarks in a 100B-token continuation experiment on Nemotron-3 Nano. The approach improved GPQA scores by +11.1 points, MMLU-Pro by +1.8, average code by +1.9, and commonsense understanding by +1.6.
Comments
Loading...