NVIDIA's ScaleEvals (CoRL 2026) scales robot evals with reset-free autonomous play, near-zero humans
gharik · x · 2026-11-09
youliangtan presents ScaleEvals, accepted to CoRL 2026 (work done at NVIDIA 6 months ago). The problem: generalist robot models are scaling but real-world evaluation isn't — humans are the bottleneck, endlessly resetting scenes, setting tasks, and grading rollouts, making evals slow and unreliable.
ScaleEvals introduces reset-free autonomous play with almost no human involvement, while maintaining high correlation with manual evaluations — making robot evaluation scalable.
More from Embodied
- Delta Intelligence shows single generalist policy driving 69-DoF humanoid at near-human speed — chris_j_paxton · 2026-11-09
- VisionSIM, a Blender-based multi-sensor robot perception simulator, lands CoRL 2026 spotlight — CSProfKGD · 2026-11-09
- Apptronik hosts private CoRL recruiting events for locomotion, dexterous manipulation and sim2real roles — KyleMorgenstein · 2026-11-09
- CoRL paper GLIDE uses LLM-edited teleop code, lifting hard-task success from 0% to 90% — anand_bhattad · 2026-11-09
- CoRL 2026 kicks off with Georgia Tech group discussing human-robot collaboration and robot learning — siddkaramcheti · 2026-11-09
- Podcast breaks down what a humanoid robot really costs, from actuators to hands — MarwaEldiwiny · 2026-11-09