MaskHarness-WAM wraps a mask-conditioned world action model with a task-level harness. For long-horizon instructions involving multiple identical objects, the harness decomposes the task into subtasks, generates and verifies a target mask from the current observation at each subtask transition, and only advances after confirming completion. On a real robot platform, the system substantially outperformed limited-horizon policies on sequential multi-object manipulation. This is interesting because it addresses a core challenge in robotics: reliably performing long tasks when objects are visually indistinguishable, which current policies often fail to handle.
Target masks extend robot policies to ordered, long-horizon manipulation
MaskHarness-WAM repeatedly identifies target instances and verifies subtask completion, allowing a real robot to manipulate similar-looking objects in sequence.
Academic
Zitai Huang · Taiyi Su · Jian Zhu · Jianjun Zhang · Chong Ma · Tianbin Liu · +3 more
Tongji University · AIRC, Midea Group
Research Digest··2 min read
Thread:Live Software Adaptation
In one line
MaskHarness-WAM enables reliable sequential manipulation of identical objects by updating instance masks at each subtask transition.
Read it as
Huang et al. introduce a harness that connects high-level task planning to short-horizon visuomotor policies through instance-specific image masks.
The authors built MaskHarness-WAM, a control harness for tasks in which a robot must manipulate multiple objects—potentially identical in appearance—in a prescribed order.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.