Discover what our robot and autonomy experiments reveal. Each card summarises reviewed public research with clear scope and limitations.
V6 Meta-World robot regression corpus
Paired policy regression study on Meta-World with scripted control families. Designed to stress reference-vs-candidate comparison, transition matrices and comparable/incomparable outcomes. Status: Research preview. Collection: Physical AI.
990 Meta-World pairs
150 goal-condition clusters
Scripted control family
Missing F2F examples noted as a methodological limitation
Note: Scripted policies only; not commercial customer models
Note: Missing F2F coverage limits some failure-to-failure analyses
Note: Simulation corpus; not a Tier-3 hardware validation study
V7 world-model assurance study
Claim-scoped scorecards for lightweight learned proxies that compare predicted outcomes against executed robot trajectories. Status: Research preview. Collection: Physical AI.
Lightweight learned proxies
Claim-scoped scorecards
Failure-blind stress model documented as a genuine limitation
Simple PPI coverage failure excluded from release authority
Note: Simple PPI undercoverage must not authorize release decisions
Note: Research methods status: NO_GO for production gates
Note: Do not claim lower RMSE but worse release decision as an observed published result
V8 simulated spacecraft control study
Autonomous mission comparison on a custom Clohessy–Wiltshire free-flyer simulator with classical PD controllers and safety-filter interventions. Status: Research preview. Collection: Space Technology.
Custom CW simulator (arisole_cw_freeflyer_v1)
100 paired mission comparisons
25 mission-condition clusters
Classical PD controllers
Note: No flight or hardware validation
Note: No NASA/Astrobee integration claim
Note: Not spaceflight certification
FAQ
Are these peer-reviewed publications?
Treat them as research previews and technical summaries unless a venue entry says otherwise. SSRN human-movement papers live on /research.
Where do I explore a real evaluation?
Open /break/benchmark-001 or the V6 card above, then continue into MotionBench.