Mind2Act World: A Diagnostic Benchmark for Embodied Mind, Physical Action, and Mind-Action Composition in Robotic Manipulation
Yuzhi Huang, Jie Wu, Fanding Huang, Junchao Lin, Weichen Zhang, Ziyi Xiong, Yang Li, Zhengwei Xie, Zhimin Wang, Zhen Liu, Jiazhen Huang, Yuzheng Zhang, Yuche Liu, Hongyi Yang, Chen Tang, Shixiang Tang, Hakan Aktas, Adam Wu, Duo Wu, Kangye Ji, Shuzhao Xie, Huanan Liu, Weijue Bu, Chenxin Li, Xinyuan Wang, Tianyi Bai, Yinan Mao, Jiawei Wang, Minglei Li, Heng Yu, Ehsan Adeli, Minchen Li, Yong-Lu Li, Cengiz Öztireli, Huazhe Xu, Chao Yu, Jing Zhang, Shanghang Zhang, Weitao Zhou, Wenwu Zhu, Cheng Chi, Jingyan Jiang, Zhi Wang
cs.RO, cs.AI, cs.CV
2026-11-20
Mind2Act World tests 30 bimanual tasks in three views: memory, execution, closed-loop composition. Best agent scores 52.7% and 39.0% on the first two, but 12.0% on composition.
Manipulation needs two distinct capabilities. Reasoning holds onto information that is no longer visible, infers rules from feedback, and keeps task state current; execution produces the required physical effect inside a time window. Most benchmarks score whole tasks end to end, so a failure cannot be traced to either side. A few split reasoning and execution into separate assessments, but none test whether the two still hold together when one realized outcome sets the conditions for the next decision. A 16-institution team led by Tsinghua turned that attribution problem into a reusable diagnostic benchmark.
Thirty bimanual tabletop tasks (two ARX X5 arms in Isaac Sim), drawn from five domains including food service, manufacturing, and instrument operation, are organized into three views of ten tasks each:
Each task has three difficulty levels and two scene settings; Clean and Random share the same task instance, and Random only adds clutter, 23 table materials and six lighting types, for 180 configurations. Scoring is environment-side: a deposit counts only after release and support, a dial must hold its reading, and self-declared completion earns nothing. Two interfaces are supported. End-to-end policies, VLAs that map images and instructions straight to actions and WAMs that predict world change before acting, are fine-tuned on 4,500 shared demonstrations (75.8 hours). Agents act through one fixed tool harness (moveeef, memoryresolve) with no training. Mind2Act demonstrations average 99.6 s, against 13 to 40 s in benchmarks like RoboTwin 2.0 and RoboDojo. A diagnosis harness combines rule-based analysis of stage and event records with an LLM judge to attribute each failed episode to Mind, Act, or their composition, then summarizes each system in a profile card. The Overall score weights composition heaviest: 0.25 Mind + 0.25 Act + 0.50 Mind2Act.
Nine systems, 56,700 scored episodes, two separate leaderboards. GPT-6 Astra leads agents at 34.6% overall success, InternW0 leads end-to-end policies at 8.7%, and the remaining seven systems sit between 2.9% and 7.0%.
| System | Type | Mind | Act | Mind2Act |
| GPT-6 Astra | agent | 52.7% | 39.0% | 12.0% |
| DeepSeek-V4.1-Flash | agent | 11.0% | 8.7% | 0.0% |
| Gemini-4-Argon | agent | 4.3% | 9.7% | 0.0% |
| InternW0 | WAM | 5.4% | 17.6% | 3.0% |
| X-VLA | VLA | 1.0% | 19.6% | 0.2% |
| π0.5 | VLA | 3.5% | 12.0% | 3.6% |
Facts that matter more than the totals:
One honest nuance: partial scores show most Mind2Act episodes die at their first steps, before reasoning and execution ever have to cooperate. G0.5 never grasps a piece in 202 of its 300 bridge episodes; the clearest true composition failure is InternW0, which switches on the heat before all ingredients are in the pot in 271 of 300 cooking episodes.
This benchmark swaps a leaderboard for a physical exam: which layer is weak, where the first breakdown lands, what the failure mode looks like, and the profile card points at the next fix. For VLA developers it flags two attackable problems: memory that evaporates under appearance change, and a 3.6% ceiling on composed tasks. The 4,500 long-horizon demonstrations are themselves a scarce asset, since comparable benchmarks average 13 to 40 s per demo. The progress here is in evaluation, not in capability; putting composition under isolated pressure is the part nobody had systematized.
The paper concedes the central one itself: the three views contain different tasks and Mind2Act workflows are longer, so view-level gaps alone do not establish a composition failure, and matched interventions are left to future work. The timing setup deserves care: simulation pauses during inference, so timed tasks measure reaction in simulation time and an agent's multi-second latency stays invisible. The two boards are not strictly comparable either, since agents run 5 episodes per configuration against 50 for policies on a 4x physics budget. The diagnosis harness covers only three representative systems, and every task is a tabletop abstraction rather than a deployed workflow.