Tag
A new blog post traces the evolution of Vision-Language-Action (VLA) models from RT-1 to MolmoAct2, explaining why VLAs are a leading bet for generalizable robot learning and covering key architectural trade-offs.
The paper introduces uni-robot-agent, a unified harness enabling open-weight robot-use agents to outperform proprietary models like GPT-6 Astra-RoboProbe on zero-shot manipulation benchmarks, and identifies four core capabilities that future robot-use agent training should jointly develop.
HiFOR proposes a hierarchical foresight framework with action-guided world retrieval for vision-language-action models, jointly predicting short-term states, semantic milestones, and goal states to better guide robotic manipulation action generation.
OpenFlyV2 是一个统一的机载视觉-语言-动作模型，在资源受限的硬件（NVIDIA Jetson Orin NX）上以 10Hz 实现实时空中导航与推理，并能从仿真零样本迁移到真实四旋翼飞行，达到 48% 的导航成功率。
Google brings Gemini Robotics Embodied Reasoning 2 (ER 2), a specialized embodied reasoning model for robots, to public preview on Google Cloud's Gemini platform for enterprise deployments, enabling spatio-temporal reasoning, long-horizon planning, and multi-robot collaboration.
The paper introduces Power-of-Two Leaky Integrate-and-Fire (PoT-LIF) neurons that emit signed, power-of-two amplitude spikes to enable event-driven spiking computation in Vision-Language-Action models without restructuring pretrained policies. Integrated into the π0.5 VLA projections, it matches or exceeds the continuous-activation baseline on LIBERO benchmarks while reducing modeled arithmetic energy by about 86.95%.
TAD introduces Trajectory-only Adversarial Distillation, which improves vision-language-action policies using only a fixed bank of teacher trajectories—no teacher action scores or task rewards—via a discriminator providing on-policy rewards, boosting OpenVLA-OFT success rates substantially while trailing white-box on-policy distillation.
PDRVLA 是一篇机器人操作研究论文，提出在视觉-语言-动作（VLA）模型中加入轻量级物理动力学模块，通过世界模型监督预测未来视觉特征，从而在不引入在线视频生成延迟的情况下提升机器人动作生成性能，并在 LIBERO、RoboTwin 2.0 等仿真基准上取得领先成功率。
ThinkWorld 是一个 VLA 学习框架，通过 Chain-of-Dynamics（CoD）运动 token 表示动作，并利用世界想象进行强化学习，从而从无标注人类视频中获取监督，提升机器人操作在未见任务和环境中的泛化能力。
论文提出 Just Action Model (JAM)，将视觉-语言-动作模型与世界-动作模型解耦为骨干初始化、未来观测目标与推理机制三个可分离选择，并用仅在训练期使用的 foresight loss 在 LIBERO、RoboTwin 等基准上超越 FastWAM/ImageWAM，同时减少 73% 的训练 FLOPs。
SO-101 Bench 推出一个基于 Isaac Lab 数字孪生的低成本机器人臂语言条件空间操作仿真基准，结果显示混合 LLM 智能体（83.2%）显著优于学习型 VLA 策略（约61%），且仿真与真实硬件上的策略排名完全一致。
HuGe 提出了一族人形通用视觉-语言-动作（VLA）模型，利用超过2000小时的第一人称人类与机器人数据进行预训练，并通过伪本体感知与视觉适配实现全身控制，在 HumanoidArena 仿真基准上达到 SOTA，同时在 Unitree G1 实体机器人上验证了其有效性。
RynnVLA-Latent 提出用 30.1K 小时异构视频学习 3D-grounded 潜在动作，预训练 4B 视觉-语言-动作策略，在 LIBERO（98.8%）、RoboTwin 2.0 及真实机器人操作任务上取得领先成功率，无需针对基准的专门调优。
This paper extends the Platonic Representation Hypothesis to robotics, introducing Sensorimotor Alignment (S²) as an offline proxy that measures overlap between a model's sensory representations and end-effector trajectories, showing it correlates with downstream robot task success across VLAs, WAMs, and real-robot policies.
Sam Levine highlights new work showing that extracting policies via supervised learning conditioned on the advantage can reach state-of-the-art RL results. Paul Zhou introduces Advantage Guidance, a stable advantage-conditioned supervised training method with contrastive guidance sampling, improving on his earlier π*0.6 VLA work at Physical Intelligence.
KPHarness introduces a framework that co-evolves recovery knowledge and active probing strategies to adapt the external execution harness of a frozen vision-language-action model for long-horizon Minecraft tasks, improving task coverage and transfer to unseen tasks without further updating the underlying actor.
论文提出 SERA（Self-Evolving Robotic Agent），通过可编辑的 harness 维护任务上下文并利用执行反馈自我进化规划，配合固定的 VLA 执行器来改进长时程机器人操作任务，在 RoboMME 上取得 51.38% 的成功率。
RoboCasa-Shift 提出一个基于 RoboCasa365 的阶段感知鲁棒性基准，通过配对 rollout 与受控扰动评估视觉-语言-动作（VLA）策略在长程家务操作任务中的表现，发现失败模式随扰动因子和任务而异，并非均匀退化，主张鲁棒性应作为执行的多维属性而非单一端到端指标衡量。
This paper presents a mechanistic interpretability study of vision-language-action (VLA) models, showing that motor programs are bound to absolute coordinates and ignore scene changes, explaining perturbation brittleness. It introduces Action Atlas, an interactive platform covering 520+ trained sparse autoencoders and 79 identified concepts across six VLA models.
VLDA提出用紧凑的Delta变量将机器人控制重构为可操作的转变建模，作为视觉语言理解与动作生成之间的接口，相比直接观测-动作映射和完整未来状态预测，提升了长程一致性、鲁棒性和泛化能力。