Data compilation interface and task translation view
Data Compilation:
Observation in. Operation out.

ArcheBase combines task reasoning, physical-causal modeling, and data engineering to compile observational robot episodes into operational data for training, evaluation, and iteration.

See the Compilation Flow

「 Why Compilation 」

A recording is not a task asset.

Observational data records what happened. Operational data preserves task state, contact, failure, recovery, and the physical evidence models need to learn what to do next.

01

Unclear task boundaries

A continuous recording does not reveal where one task stage ends and the next begins.

02

Hidden intent

Motion alone does not explain the goal, the operator's judgment, or the reason for recovery.

03

Unaligned signals

Video, motion, force, and robot state arrive on different clocks and in different formats.

04

Unstructured failure

Without task context, deviations, errors, and recovery remain noise instead of reusable evidence.

「 Physical Data Compiler 」

A system that turns observation into operation.

Physical AI DBMS organizes and traces task episodes. The Physical Causal Model interprets task state, contact, failure, recovery, and outcome.

Input layer

Observational input

Video, trajectories, sensor logs, environment states, and outcome records show what happened.

Human operationTask objectiveEnvironment stateVisual & sensor streamRobot feedback

Compilation layer

Physical AI DBMS

A database system for synchronized multimodal episodes, task indexing, versioning, lineage, quality, and agent-native retrieval.

Time synchronizationMultimodal indexingTask slicingQuality verificationVersion & lineageAgent-native query

Reasoning layer

Physical Causal Model

A Physical Causal Model that reconstructs task stages, state transitions, contact, failure causes, recovery paths, and outcomes.

Task-state reasoningContact dynamicsFailure attributionRecovery pathsOutcome evidence

Output layer

Operational output

Compiled assets capture task stages, contact, failure causes, recovery, and quality signals.

Training datasetBenchmarkScene data packageEvaluation assetsReusable task templates

Input layer

Observational input

Video, trajectories, sensor logs, environment states, and outcome records show what happened.

Human operationTask objectiveEnvironment stateVisual & sensor streamRobot feedback

Compilation layer

Physical AI DBMS

A database system for synchronized multimodal episodes, task indexing, versioning, lineage, quality, and agent-native retrieval.

Time synchronizationMultimodal indexingTask slicingQuality verificationVersion & lineageAgent-native query

Reasoning layer

Physical Causal Model

A Physical Causal Model that reconstructs task stages, state transitions, contact, failure causes, recovery paths, and outcomes.

Task-state reasoningContact dynamicsFailure attributionRecovery pathsOutcome evidence

Output layer

Operational output

Compiled assets capture task stages, contact, failure causes, recovery, and quality signals.

Training datasetBenchmarkScene data packageEvaluation assetsReusable task templates

Input layer

Observational input

Video, trajectories, sensor logs, environment states, and outcome records show what happened.

Human operationTask objectiveEnvironment stateVisual & sensor streamRobot feedback

Compilation layer

Physical AI DBMS

A database system for synchronized multimodal episodes, task indexing, versioning, lineage, quality, and agent-native retrieval.

Time synchronizationMultimodal indexingTask slicingQuality verificationVersion & lineageAgent-native query

Reasoning layer

Physical Causal Model

A Physical Causal Model that reconstructs task stages, state transitions, contact, failure causes, recovery paths, and outcomes.

Task-state reasoningContact dynamicsFailure attributionRecovery pathsOutcome evidence

Output layer

Operational output

Compiled assets capture task stages, contact, failure causes, recovery, and quality signals.

Training datasetBenchmarkScene data packageEvaluation assetsReusable task templates

Input layer

Observational input

Video, trajectories, sensor logs, environment states, and outcome records show what happened.

Human operationTask objectiveEnvironment stateVisual & sensor streamRobot feedback

Compilation layer

Physical AI DBMS

A database system for synchronized multimodal episodes, task indexing, versioning, lineage, quality, and agent-native retrieval.

Time synchronizationMultimodal indexingTask slicingQuality verificationVersion & lineageAgent-native query

Reasoning layer

Physical Causal Model

A Physical Causal Model that reconstructs task stages, state transitions, contact, failure causes, recovery paths, and outcomes.

Task-state reasoningContact dynamicsFailure attributionRecovery pathsOutcome evidence

Output layer

Operational output

Compiled assets capture task stages, contact, failure causes, recovery, and quality signals.

Training datasetBenchmarkScene data packageEvaluation assetsReusable task templates

Input layer

Observational input

Video, trajectories, sensor logs, environment states, and outcome records show what happened.

Human operationTask objectiveEnvironment stateVisual & sensor streamRobot feedback

Compilation layer

Physical AI DBMS

A database system for synchronized multimodal episodes, task indexing, versioning, lineage, quality, and agent-native retrieval.

Time synchronizationMultimodal indexingTask slicingQuality verificationVersion & lineageAgent-native query

Reasoning layer

Physical Causal Model

A Physical Causal Model that reconstructs task stages, state transitions, contact, failure causes, recovery paths, and outcomes.

Task-state reasoningContact dynamicsFailure attributionRecovery pathsOutcome evidence

Output layer

Operational output

Compiled assets capture task stages, contact, failure causes, recovery, and quality signals.

Training datasetBenchmarkScene data packageEvaluation assetsReusable task templates
Explore Physical AI DBMSExplore the Physical Causal Model

「 Compilation Workflow 」

How a task episode becomes an asset.

The workflow adds task structure, physical causality, quality validation, and versioned delivery without reducing a real task to a loose set of files.

The workflow adds task structure, physical causality, quality validation, and versioned delivery without reducing a real task to a loose set of files.

「 Deliverables 」

Operational data assets, ready for training and evaluation.

The delivery is not a loose file set. Each task asset supplies aligned evidence for training, benchmarked evaluation, failure analysis, and repeated task use.

The delivery is not a loose file set. Each task asset supplies aligned evidence for training, benchmarked evaluation, failure analysis, and repeated task use.