「 Inside the Physical Data Compiler 」
Physical Causal Model
compiles physical evidence into operational data.
Language reasoning interprets task context. Physical causal inference grounds state transitions, contact, failure, recovery, and outcome in synchronized real-world evidence, producing data that models can train on and evaluate against.
Explore physical causal compilation「 Why A Physical Causal Model 」
Knowing what appeared is not the same as knowing what caused the result.
Physical AI must connect actions to state changes, contact to outcomes, and failure to recovery. Language priors help reason; physical evidence keeps the reasoning grounded.
01
Tasks unfold through state
A task is a sequence of changing conditions, not a single label attached to a video.
02
Contact changes what is possible
Forces, constraints, and object interaction determine whether an action succeeds, fails, or requires recovery.
03
Failure carries causal information
Deviation and recovery reveal task boundaries, hidden constraints, and the evidence needed for evaluation.
「 Physical Causal Compilation 」
Operational data is produced from evidence, not labels alone.
The compiler reconstructs task state, identifies contact and constraints, validates causal transitions, and packages the result as trainable, evaluable, and queryable operational data.
「 Compiled Operational Data 」
Operational data assets for training, evaluation, and iteration.
The same physical evidence can be compiled into task schemas, failure and recovery cases, and causality-aware evaluation assets, then accessed through Data APIs.
Task-stage reconstruction
Recover approach, alignment, contact, execution, failure, recovery, and completion from continuous episodes.
Failure cases
Package deviations with task state, physical constraints, contact evidence, causal attribution, and outcome.
Recovery cases
Package comparable recoveries and the state transitions that returned a task to a valid path.
Causality-aware evaluation
Generate evaluation cases, criteria, and regression slices grounded in real outcomes and evidence.