Research

Understanding why
physical systems fail.

We study causal structure, model failure, and the experience required for machines to generalize in the physical world.

Illustrative white robotic arm reaching toward a cube in a controlled lab
Original Friction: Low
Lighting: Normal
Intervention A Friction: High
Lighting: Normal
Intervention B Friction: Low
Lighting: Dim
Our research areas

Causal Understanding

Discovering cause–effect relationships in complex real-world environments through controlled interventions and rigorous evaluation.

Learn More

Causal Discovery

Identify factors that truly influence outcomes.

Controlled Experiments

Isolate variables using structured interventions.

Failure Analysis

Explain model failures through causal attribution.

Generalization

Build models that transfer across real-world shifts.

Featured research

Robotic manipulation experiment Benchmarking Physical AI
with Causal Perturbations Robotic workcell with causal model trajectories Causal World Models
for Robotics Robotic workcell with diverse cube experiments Experience-Guided
Data Generation Measuring Generalization
Beyond i.i.d.

Our methodology

How we study failure.

Our evidence-first workflow turns observations into causal insight and systematic improvement.

Learn More About CPF

Failure Evaluation

Detect real-world or simulated failures.

Controlled Intervention

Systematically perturb variables to test causal impact.

Failure Attribution

Identify root causes through causal analysis.

DataGap & DataPlan

Decide which experience is missing and what to collect next.

Validation

Collect, retrain, and validate for real improvement.

Evidence-first closed-loop research workflow.

Current research status

Current foundation

  • External technical overview
  • Phase 1 study protocol
  • Evidence-first validation boundary

In Progress

  • Episode Bundle verification
  • Retraining validation studies
  • Data-plan comparison runs

Planned

  • Broader simulator benchmarks
  • Cross-model causal studies
  • Design partner validation

Technical resources

DOCCPF Technical Overview

External, evidence-first overview of the Causal Perturbation Framework.

External version DOCPhase 1 Experiment Protocol

Controlled study design for intervention and data-collection comparisons.

Protocol IN DESIGNCPF Technical Demo

Evidence-first demonstration storyboard and validation path.

Validation pending DOCJoint Validation Plan

External brief for a small-scale, evidence-first validation program.

External brief

Interested in joint validation
or research collaboration?

We are seeking conversations with robotics companies, AI labs, and physical AI teams on evidence-first joint validation.

Contact Research Team View Technical Overview
Research ecosystem & collaboration Robotics LabsOpen ResearchSimulation ToolsPhysical AI TeamsDesign Partners