Event

Research team releases Embodied Playground long-horizon embodied reasoning benchmark

1 reports · 1 sources

Feng Chen and other researchers propose and release Embodied Playground, a benchmark for evaluating how foundation model agents coordinate physical, strategic, and spatial reasoning in long-horizon tasks. The benchmark contains 410 task environments, covering basic manipulation, sequential puzzles, and multi-stage household activities, with differences set in planning depth, geometric constraints, and physical execution requirements. The researchers curated EP-100 from it: 100 representative tasks across five suites and 16 task types, while also providing EP-Easy and EP-Hard difficulty splits and a streamlined EP-20 version. Evaluation uses a unified protocol that specifies observation and tool access, resource budgets, and repeated trials, and separately checks goal completion, task rule compliance, and physical validity.