Skip to content
Only Research#Software and compute systems
  1. arXiv cs.RO79

    A2-MEAS: An Embodied Multi-Agent Framework with Code as a Scaling Interface

    The authors propose A2-MEAS, an embodied multi-agent framework that uses code as a scaling interface, combining corrective language, spatial evidence, execution history, and action proposals into correction codes; during training, MetaBrain provides supervision and runtime guidance.

    Why it matters: The framework uses code as a scaling interface, injects corrective language and execution history into the policy, and provides quantitative comparisons across benchmarks and real-robot tasks.

  1. 3D视觉工坊 · 微信公众号88

    Infinigence AI, Tsinghua and AgiBot propose Collective Experience Learning (CEL), validated on 96 robots with 98.3% success rate

    Infinigence AI, together with Tsinghua University and AgiBot, proposed the Collective Experience Learning (CEL) paradigm and completed prototype validation on 96 robots. In a phone packaging task, policies trained with CEL completed 120 tests on two robots of the same model that did not participate in training, with the success rate rising from 79.2% to 98.3%; when the number of robots participating in learning increased from 16 to 96, the test success rate rose from 86.7% to 98.3%.

    Why it matters: Fault-tolerant asynchronous training across a 96-robot real-machine cluster and failure-guided experience retrieval provide a reproducible system and scale-benefit data for Collective Experience Learning.

  1. CVer · 微信公众号77

    Tsinghua AIR proposes OxyGen: unified KV Cache management lets robots talk while acting

    The Institute for AI Industry Research (AIR) at Tsinghua University, together with Tsinghua's Department of Electronic Engineering and the University of Science and Technology of China, has proposed OxyGen, which uses unified KV Cache management to let an embodied model's action and language reasoning share prefix computation. The paper has been accepted to NeurIPS 2026.

    Why it matters: The paper presents a unified KV Cache management scheme and measured data across multiple platforms, offering a reference for on-device embodied inference deployment.

  1. arXiv cs.RO78

    Study: Full-weight VLA fine-tuning without an fp32 master copy silently erases most updates

    The paper points out that when VLA models are fully fine-tuned in bf16 without keeping an fp32 master copy, AdamW rounds away any update smaller than 1/512 of a weight; 9 of 12 open-source VLA codebases train this way by default or in documented recipes, including LeRobot's SmolVLA default configuration and the π0/π0.5 guides.

    Why it matters: The original quantifies the proportion of updates lost in bf16 fine-tuning and provides a LIBERO success-rate comparison for the fp32 master copy, which can serve as a reference for training configurations.

You’re all caught up