Chen FengEmail
← All publications

Embodied Playground: A Long-Horizon Agentic Robotics Benchmark for Physical, Strategic, and Spatial Reasoning

Feng Chen†*, Pei Zhou*, Lixuan He*, Yunchao Yao*, Li Sun*, Shenghua Gao, Mingyu Ding, Kaifeng Zhang, Kaichen Zhou, Yi Ma

† Project leader; * Core contributor.

arXiv2026Agentic Robotics

Embodied Playground contains 410 simulated environments. In a Tower of Hanoi task, for example, the robot has to plan a sequence of moves and follow the stacking rules at every step. Reaching the target arrangement only counts if the moves that got it there were legal.

For a smaller evaluation, we selected 100 tasks across five suites for EP-100. Each task has four starting conditions. The protocol specifies what an agent can observe and which tools it can use, and checks its actions as well as the final state.

The benchmark code is available on GitHub.

TAKE A LITTLE BREAK

Dino Run

ONE DINO. ONE MORE TRY.
SCORE 00000BEST 00000

Ready to run?

A tiny adventure between bigger quests.

Space / ↑ jump · P pause · Esc close
On mobile, tap the track or use Jump. Best score stays on this device.