Xolver Research

Research for bounded, reviewable physical intelligence.

We study how robots can learn and act while preserving safety checks, reproducible evidence, and meaningful operator control.

Safe Play Pretraining

Learning reusable manipulation abilities through simulated play and reviewable safety signals, with a first downstream study in contact-rich peg insertion.

Manipulation priorsSafe reinforcement learningEvidence provenance

Read the research

ContractBench-DROID Lite

Evaluating whether robot policy data carries the safety context, replay records, and evidence needed before physical deployment.

ReadinessDataset contextReplayability

Read the research

A consistent research boundary.

Across projects, learned models propose actions, review boundaries constrain them, and evidence records preserve how an artifact was produced and evaluated.