Resetting a robot's test scene by hand is slow, and now there's a system that does it for you.
A team built HALTER, a harness that automates the reset-and-score cycle for testing robot manipulation policies on long, multi-step tasks. Instead of a person resetting the workspace after every trial, or a single learned policy trying to cover every way a task could end, HALTER plans a sequence of smaller "atomic" reset skills to put the scene back. It builds a 3D scene graph from cameras and point clouds, then hands that graph to an LLM, which scores how much of the task got done, plans the reset, and checks that the reset actually worked, without collecting labeled training images. On four long-horizon tasks with a Franka arm, it restored the scene correctly 76% of the time, versus 52% for the prior system AutoEval and 65% for a motion-planning baseline. It also cut the human labor of running an eval campaign by 72%.
Real-robot benchmarks are only as trustworthy as the resets between trials, and manual resets make results hard to reproduce across labs. Because HALTER composes reset skills instead of learning one policy per end state, it generalized to three tasks it never saw during training, resetting 74.7% of episodes versus 1.3% for a task-specific reset policy, a sign the approach scales better than task-by-task engineering.
It is a plumbing fix, not a flashy one, but plumbing is what has been missing from robotics benchmarks that claim to measure real progress.