Robot Data and Hardware

Design robot policy uncertainty and abstention with calibration, risk-coverage curves, lead time, human handoff, fallback outcomes and safe re-entry.

Choose LoRA, adapters or full robot VLA fine-tuning by distribution shift, data and compute. Test retention, action decoding and deployment bundles.

Design robot policy runtime safety shields with executable constraints, CBFs, predictive filters, state uncertainty, deadlines, fallbacks and audits.

Augment robot vision data without breaking action labels. Audit crops, flips, camera geometry, temporal consistency, clipping and hardware replay.

Version robot datasets with immutable IDs and manifests linking raw episodes, labels, transforms, splits, statistics, training runs and deployed models.

Prevent robot evaluation leakage with episode, session, site, object and operator group splits, sibling detection, train-only preprocessing and frozen tests.

Mine robot failures from rollout windows using uncertainty, novelty, impact, clustering and human review, then recollect safely and retest holdouts.

Collect human-in-the-loop robot intervention data with control authority, pre-takeover context, correction intent, timing, safe handback and evaluation.

Weight robot dataset mixtures by domain and outcome while controlling sampling exposure, oversampling, episode length, drift and worst-domain performance.

Normalize robot actions across embodiments while preserving command type, frames, units, gripper meaning, timing, masks, statistics and round-trip tests.