EvoAction

Annotating videos for temporal action segmentation is expensive and time-consuming, especially for Physical AI, where identifying precise transitions between robotic subtasks is critical for training models. EvoAction is a recursively self-improving agentic harness for data annotation focused on temporal action segmentation. Built on Crusoe Managed Inference with NVIDIA Nemotron, it autonomously annotates videos, evaluates failures, and spawns specialist agents to improve its own workflows, prompts, and tools. Each new version is independently evaluated, with successful improvements retained and regressions rolled back. We use WGO-Bench to measure annotation quality and evaluate improvements on unseen robot demonstrations.