2026-11-13 15:49

Introduction: 1.2 million hours of real-world experience, targeting zero-shot generalization in robotics. What is Physis-v1.0-preview? Physis-v1.0-preview

What is Physis-v1.0-preview?

Physis-v1.0-preview is the pre-release version of the universal physical intelligence foundation model launched by Inverse Matrix (Physis). Centered on real-world dynamics prediction under action conditioning as its core pre-training task, Physis-v1.0-preview learns physical laws from 1.2 million hours and 148 million segments of multi-source real-world experience. It is accompanied by a full-stack infrastructure covering data factory, efficient training (MFU exceeding 40%), and real-device deployment, enabling cross-scenario transferability and zero-shot generalization.

Key Features of Physis-v1.0-preview

• Physical World Understanding: Learns representations of objects, scenes, and physical relationships from real-world video and operation data, forming an understanding of dynamic evolution.
• Action-Conditional Prediction: Given current observations and candidate actions, predicts state changes in the world, providing "imagination" capabilities for decision-making.
• Cross-Scenario Generalization: Reuses existing knowledge to perform zero-shot execution in unfamiliar environments, new objects, or novel tasks, or rapidly adapts with minimal demonstrations.
• Real-Device Operation: Deploys model decisions onto actual robots to complete physical tasks under perception noise, contact dynamics, and actuation constraints.
• High-Frequency Closed-Loop Control: Integrates low-latency inference engines to support real-time action correction based on continuous feedback, maintaining stable control rhythm.
• Multi-Skill Composition: Decomposes high-level goals via low-frequency semantic planning and invokes skills, while high-frequency control handles execution, enabling collaborative completion of complex long-horizon tasks.

Technical Principles of Physis-v1.0-preview

• Foundation Task Design: The team selects "real-world dynamics prediction under action conditioning" as the pre-training objective. Physical interactions inherently provide supervision signals—actual world changes after action execution are recorded in subsequent observations. Despite task diversity, shared physical relationships such as contact and constraints govern state evolution, allowing heterogeneous multi-source data to provide supervision for a single transferable capability, continuously learning more general laws as data and compute scale.
• Data Engineering: Built Physis Data Factory, integrating model operators, task orchestration, distributed execution, and quality validation into a unified production system. Raw data—including general videos, first-person videos, simulation, and real-robot captures—are processed using Lance, Ray, Prefect, and other infrastructures for cleaning, fine-grained temporal captioning, geometric reconstruction, camera trajectory estimation, optical flow, and point trajectory extraction, with precise synchronization across video, 3D trajectories, and language annotations. Counterfactual paired samples are explicitly constructed to establish action-consequence correspondences, delivering final training-ready fixed data versions.
• Training Infrastructure: Employs a three-tier heterogeneous cluster streaming architecture—Raw data → Sample → Token → Model—using steady token consumption rate from training clusters to reverse-plan capacity across layers, absorbing asynchronous fluctuations via caching, queuing, and backpressure. Training combines full-shard data parallelism and context parallelism; data loaders estimate sample computational cost, attention planners partition work based on token visibility scope, eliminating synchronization waits. Self-developed attention operators support fine-grained visibility with bidirectional chunk-wise and time-causal inter-chunk relationships, improving efficiency by 30–40% over FA4; combined with unified activation value management to reduce recomputation, achieving overall MFU exceeding 40%.

How to Use Physis-v1.0-preview

The official release address for the model weights/API has not yet been disclosed.

Core Advantages of Physis-v1.0-preview

• Generalized Physical Intelligence: Focused on reusing experience and minimal adaptation, targeting zero-shot generalization for unfamiliar scenarios and new tasks.
• Unified Pre-Training Objective: Uses action-conditioned real-world dynamics prediction, enabling multi-source heterogeneous data to supervise a single transferable capability, aligning with Scaling Law principles.
• Ultra-Large-Scale Data Supply: Leverages 1.2 million hours and 148 million segments of multi-source experience through Physis Data Factory, enabling automated processing and stable delivery at the million-hour scale.
• Efficient Training System: Three-tier streaming architecture with load balancing, self-developed attention operators, and activation value management achieves MFU > 40%, ensuring high compute utilization.
• Fine-Grained Physical Understanding: Explicitly learns causal relationships between actions and consequences through fine-grained temporal annotation, geometric reconstruction, and counterfactual paired samples.
• Full-Stack Co-Design: Data, pre-training, inference, and deployment are jointly optimized around the same learning task—local optimizations directly translate into holistic performance gains.

Competitive Comparison: Physis-v1.0-preview vs. Alternatives

Application Scenarios for Physis-v1.0-preview

• Industrial Manufacturing: Fine-grained operations such as component grasping, assembly, sorting, and quality inspection, adapting to frequent line changeovers and long-tail parts.
• Logistics & Warehousing: Package picking, palletizing, depalletizing, loading, handling massive SKUs and dynamic warehouse layouts.
• Home Services: Open-ended household tasks like organizing items, cleaning, folding laundry, meal preparation—rapidly adapting to unfamiliar environments.
• Commercial Services: Robot services in unstructured settings such as meal plating, retail restocking, and hotel item delivery.
• Research & Education: Serves as a physical intelligence foundation, providing pre-trained models and full-stack infrastructure support for embodied intelligence research.

Source: AI Tools Hub