HALTER Automates Long-Horizon Robot Manipulation Evaluation via Scene Graphs
September 18, 2026
HALTER automates robot policy evaluation by using a library of learned atomic reset skills to restore scenes after long-horizon rollouts. It constructs spatial scene graphs from point clouds and vision foundation models, using an LLM to reason over the graph for automated scoring.
HOW THIS AFFECTS YOU
●
builderThis framework enables more robust benchmarking for autonomous manipulation policies in complex environments.
●
researcherYou can scale robot evaluation experiments without the bottleneck of manual scene resets.