Insight
Research
Benchmarks, notes, and research directions on data curation and real-world manipulation.
BenchmarkJan 2026
Benchmarking Visual Temporal Progress for Data Curation
OpenGVL: a benchmark for estimating task progress across diverse manipulation tasks, for both robotic and human embodiments.
Research noteFeb 2026
From the noise to the useful signal
Quality metrics that filter noise from real-world robot data to improve final policy performance.
Research direction2026
Learning manipulation from human demonstrations
How our robots learn from human demonstrations, just as humans learn from one another, and improve over time.
Research direction2026
Reward models and temporal progress across embodiments
Ranking trajectories by quality across hardware and operators to curate training data at scale.
Research direction2026
Vision-language-action models for real-world deployment
VLA modeling for manipulation tasks that have to work outside the lab, in retail, manufacturing, and logistics.