arXiv AI

Learning the Task or Learning the Verifier? Rethinking Criteria Exposure in Embodied Self-Improvement

The paper investigates the impact of exposing self‑improving embodied agents to benchmark success criteria, a practice termed Criteria Exposure (CE). Experiments show that while CE leads to higher success rates on the exposed criteria, the resulting code overfits and performs poorly when the criteria are slightly altered or when transferred to new tasks. The authors propose instruction‑faithful criteria that more accurately reflect task completion by agent‑generated programs.

arXiv AI
3d ago

EmbodiedRSI: Active Continual Robot Learning Through Hypothesis-Guided Co-Evolution

EmbodiedRSI is a self‑evolving agentic harness that autonomously selects physical experiments to refine robot code and skills, using a Fast‑Slow Dual‑System Architecture and a Hypothesis Graph. It employs Value‑of‑Information Experiment Selection to distinguish competing hypotheses, guiding Code‑Skill Co‑Evolution, while the Slow System builds Hierarchical Memory and Reward‑Grounded Memory Learning to improve future Fast‑System performance. On RoboCasa365 and LIBERO‑Pro benchmarks, EmbodiedRSI achieves markedly higher success rates than baselines and demonstrates zero‑shot transfer to real‑world robots, achieving 71.3% overall success across multiple challenging tasks.

By Python Song, Zhixuan Liang, Kelsey Fu, Mengdi Wang, Junfeng Yang, Shilong Liu
arXiv AI
Jun 19

Playful Agentic Robot Learning

arXiv:2606. 19419v1 Announce Type: cross Abstract: Current agentic robot systems can write executable Code-as-Policy programs, observe feedback, and revise behavior across multiple attempts, but they remain largely task-driven: reusable skills are acquired only after explicit instructions.

By Junyi Zhang, Jiaxin Ge, Hanjun Yoo, Letian Fu, Zihan Yang, Yaowei Liu, Raj Saravanan, Shaofeng Yin, Justin Yu, Dantong Niu, Zirui Wang, Roei Herzig, Ken Goldberg, Yutong Bai, David M. Chan, Ion Stoica, Angjoo Kanazawa, Jiahui Lei, Haiwen Feng, Trevor Darrell
Hugging Face Trending Papers
Nov 9

Embodied First Exam

The paper introduces the Embodied First Exam, a comprehensive evaluation framework for coding agents to control robots across diverse tasks and settings. It includes 1,156 tasks on 47 robots, spanning new domains such as spatial puzzles and multi-robot interaction, and offers three evaluation modes: Standard, Realtime, and Privileged. In the Standard setting, the best agent (GPT‑6 Astra in Codex) achieves only 34% success, highlighting significant room for improvement.

arXiv Computer Vision
Nov 9

Flex-RSI: Benchmarking Recursive Self-Improvement of Frontier Agents

Flex‑RSI is a new benchmark that evaluates how agents can recursively improve their own harness—prompts, tools, memory, and control loop—while solving vision and robotics tasks under a fixed budget. Each round alternates between an executor that refines the task solution and an improver that revises the harness, with only higher‑scoring solutions retained for the next iteration. The benchmark tests both closed‑source and open‑weight models across perception, reasoning, robot action, 3D reconstruction, visual creation, and infrastructure tasks, revealing that self‑improvement is most effective when agents build their own solutions, that simple success/failure feedback limits open‑weight models, and that harness editing itself contributes significantly to performance gains.

By Jingdong Zhang, Jionghao Wang, Zhengming Yu, Yancheng Cai, Wei Chow, Chris Choy, Zhiyang Dou, Hang Hua, Jiachen Li, Zhixuan Liu, Chen Wang, Songlin Wei, Siyuan Yang, Yu Yuan, Xiaohang Zhan, Yucheng Zhang, Shuhong Zheng, Xin Li, Lingjie Liu, Wojciech Matusik, Zhengzhong Tu, Wenping Wang
arXiv Computer Vision
Oct 28

SO-101 Bench: Accessible Spatial Manipulation

SO-101 Bench is a simulated benchmark for language‑conditioned spatial manipulation on the low‑cost SO‑101 robotic arm, built using an Isaac Lab digital twin of a real setup. It features four tasks across five families, 138 objects with seen/unseen splits, 916 evaluation episodes, and 3,629 simulated demonstrations for policy fine‑tuning. The benchmark’s rankings mirror real‑world performance (Spearman ρ = 1.00) and reveals that while learned policies reach about 61% success, an LLM‑driven agent that writes code and delegates grasping achieves 83.2%.

By Truman Hickok
arXiv AI
Oct 12

EmbodiedCLBench: Evaluating Continual Learning for Self-Evolving Embodied Agents

EmbodiedCLBench is the first benchmark designed to evaluate continual learning in embodied simulation environments, focusing on self‑evolving agents. The benchmark uses compositional generalization to test whether agents can learn basic tasks and then recombine those skills to solve more complex, unseen tasks. Experiments across various agent harnesses and learning methods reveal that while basic task experience improves performance on advanced tasks, this benefit does not scale with more learning samples and current agents fail to extract additional useful information.

By Jie Huang, Yanan Chen, Ruixun Liu, Kaichen He, Xinyi He, Rui Huang, Zilong Zheng, Zhenliang Zhang, Yiwu Zhong
arXiv AI
Aug 19

Teach and Grow: An Agent-Centered Architecture for General Robot Learning

Teach-and-Grow Learning (TGL) is an agent-centered architecture that transforms a few successful demonstrations into reusable Skill Blocks, enabling a robot to compose, execute, and revise behaviors in new scenes without task-specific policy retraining. The system maintains a Skill Library and structured Experience Memory to capture successes, failures, and repairs, allowing persistent reuse and agent-directed adaptation. Evaluation on the LIBERO benchmark shows state-of-the-art performance, and the authors propose a scaling-law hypothesis suggesting that accumulated reusable experience reduces future-task error and teaching demand following a power-law trend.

By Chang Nie, Zhe Liu, Hesheng Wang