Skip to content
Changda Tian

Robotics Digest

Robotics Paper Digest — 2026-07-28

5 papers

🤖 Scanned 50 new arXiv papers (cs.RO / cs.SY / cs.LG, last 48 h), picked 5 for modular & legged robotics — summarized by DeepSeek.
🤖 扫描了近 48 小时 arXiv(cs.RO / cs.SY / cs.LG)的 50 篇新论文,围绕模块化与足式机器人精选 5 篇 — 由 DeepSeek 生成双语摘要。

1. WARL: Wrench-Augmented Reinforcement Learning for Task-Agnostic Learning in Legged Robots

WARL: 基于力觉增强的强化学习在腿式机器人任务无关学习中的应用

Figure from 2607.24036

Authors / 作者: Keita Yoneda, Kento Kawaharazuka, Kei Okada
arXiv: 2607.24036 · PDF

Proposes Wrench-Augmented Reinforcement Learning (WARL) that adds force-torque to the action space to expand exploration. Combines wrench-guided exploration with a success-rate-based switching curriculum. Demonstrated on a quadruped robot achieving robust locomotion across diverse terrains without task-specific reward tuning.

中文摘要: 提出力觉增强强化学习(WARL),将力和力矩引入动作空间以扩展探索。结合力觉引导的探索和基于成功率的切换课程机制。在四足机器人上实验,WARL能在多种地形和任务中鲁棒学习,无需针对特定任务调整奖励或设计复杂课程。消融实验验证了逐渐消除力觉的切换课程的有效性。同时指出引入力觉可能鼓励不充分利用机器人物理结构的动作,设计与机器人物理结构一致的探索增强方法仍是未来关键挑战。

💬 Novel integration of wrench in action space for RL on legged robots, improving exploration and robustness.
💬 在腿式机器人强化学习中创新性地在动作空间中引入力觉,提升了探索能力和鲁棒性。

Why read it / 推荐理由: Directly addresses exploration in legged robot RL with real-robot results on diverse terrains. 直接解决了腿式机器人强化学习中的探索问题,并在多种地形上获得真实机器人实验结果。


2. Learning Reusable Hybrid Motion Priors for Humanoid Locomotion from Motion Imitation

从运动模仿学习可重用混合运动先验用于人形机器人运动

Authors / 作者: Valerio Belli, Valerio Modugno, Enrico Mingo Hoffman, Fabio Amadio
arXiv: 2607.24083 · PDF

Three-stage pipeline: train an expert to imitate retargeted human motion, distill into a frozen hybrid motion prior (HMP) with a vector-quantized codebook, then train task-level policies by selecting discrete codebook entries. Evaluated on velocity tracking and fall recovery in simulation, and deployed on a real Unitree G1 robot.

中文摘要: 提出三阶段流程:首先训练专家策略模仿重定向的人体运动捕捉片段,然后通过蒸馏得到由本体感知编码器、残差矢量量化码本和动作解码器组成的冻结混合运动先验(HMP),最后训练任务级策略通过选择离散码本条目解决运动任务。在仿真中评估速度跟踪、点目标导航和跌倒恢复速度跟踪,并在真实Unitree G1机器人上部署速度跟踪策略。蒸馏过程保留了专家跟踪行为,HMP可无需重训练直接用于不同下游运动策略。

💬 Provides a reusable motion prior framework for locomotion, enabling fast adaptation to new tasks without retraining the base policy.
💬 提供了一种可重用的运动先验框架,实现了运动技能的跨任务复用,并进行了真实机器人验证。

Why read it / 推荐理由: Demonstrates a method to reuse learned locomotion skills across tasks, with real-robot validation. 展示了如何将学习到的运动技能在不同任务间复用,加速了新任务的训练,具有真实机器人部署验证。


3. Learning Traversability-Aware Global Planners for Long Horizon Off-Road Navigation

学习可通行性感知全局规划器用于长距离越野导航

Figure from 2607.23743

Authors / 作者: Kasi Viswanath, Jason M. Gregory, Shaunak Kolhe, Srikanth Saripalli
arXiv: 2607.23743 · PDF

Learns continuous traversability maps from overhead data (satellite, aerial LiDAR, vector maps) supervised by human driving trajectories and shaped by self-supervised geometric priors. Field trials on a Clearpath Warthog achieve path lengths within 3.66% of human-driven routes and reduce operator interventions by ~85% compared to local-planner-only autonomy.

中文摘要: 提出从头顶数据(卫星、机载LiDAR和矢量地图)学习连续可通行性地图的方法,由人类驾驶GPS轨迹监督,并通过自监督几何先验从LiDAR塑造。同时发布包含299个场景、约1244平方公里和1130公里人类驾驶数据的公共数据集。在Clearpath Warthog上跨两个地点七条路线的现场试验中,该方法达到人类路径长度的3.66%以内,并将操作员干预减少约85%。

💬 Relevant for path planning for legged robots in rough terrain, using learned traversability from aerial data.
💬 针对越野环境的长距离导航,利用学习到的可通行性地图进行全局规划,对腿式机器人的路径规划有参考价值。

Why read it / 推荐理由: Addresses long-horizon navigation in outdoor environments, which is important for autonomous legged robots. 解决了户外长距离导航中的可通行性感知问题,是腿式机器人自主导航的重要组成部分。


4. Effective Parameters, Real Behavior: Renormalization for Robotics — From Infinite Electron Mass to Sim-to-Real Gap

有效参数,真实行为:机器人学的重整化——从无穷电子质量到仿真到现实差距

Figure from 2607.24079

Authors / 作者: Youran Sun, Jiaxuan Guo, Xingyu Ren, Chugang Yi, Haizhao Yang
arXiv: 2607.24079 · PDF

Proposes a renormalization approach for sim-to-real: use effective, resolution-dependent parameters to absorb simulator omissions and reproduce real behavior. Demonstrates mechanism on PD control, dynamic rope manipulation, and underwater swimming. Provides a practical procedure for choosing observables and identifying omitted physics.

中文摘要: 提出利用重整化方法弥合仿真到现实差距:使用依赖于分辨率的有效参数来吸收仿真器省略的细节,从而再现真实行为。这些参数可能与实测物理值不同,因为它们补偿了仿真器的缺失。在有限仿真频率下的PD控制中分析其机制,展示了动态绳子操作和水下游泳的例子,并给出了选择观测变量、识别省略物理和确定有效参数的实际步骤。

💬 Offers a complementary theoretical perspective on sim-to-real transfer, which is crucial for robotic applications.
💬 从理论角度提出了一种弥合仿真到现实差距的补充方法,无需高保真仿真。

Why read it / 推荐理由: Presents a principled method to bridge the sim-to-real gap without requiring high-fidelity simulation. 提供了一种系统性的理论方法来处理仿真到现实迁移问题,对腿式机器人实际应用具有重要意义。


5. Bridging Reinforcement Learning and Optimal Control via Feasible Action Mapping

通过可行动作映射桥接强化学习与最优控制

Figure from 2607.23930

Authors / 作者: Stefan Richter, Alberto Giammarino, Guillem Torrente, Sam Blakeman, Peter Dürr
arXiv: 2607.23930 · PDF

Introduces FAOC, a framework that maps RL actions to feasible parameters of an optimal control problem, ensuring constraint satisfaction. Combines the safety of OC with the flexibility of RL. Applied to robot table tennis, outperforming baselines in sample efficiency and closed-loop performance.

中文摘要: 提出可行动作最优控制(FAOC),将强化学习与最优控制结合。关键贡献是计算高效的基于优化的映射算法,将强化学习智能体的动作从静态抽象集转换为状态依赖的最优控制问题可行参数集,严格保证动态系统约束满足。FAOC有效结合了最优控制的可预测安全性与强化学习的灵活性。在机器人乒乓球实时运动规划中,FAOC在样本效率和闭环性能上均优于当前最优基线。

💬 Combines RL flexibility with OC safety guarantees, which is useful for constrained locomotion tasks.
💬 将强化学习的灵活性与最优控制的约束满足相结合,适用于需要保证安全性的控制任务。

Why read it / 推荐理由: Provides a method to integrate RL with model-based control, enhancing safety and efficiency in complex tasks. 提供了一种结合强化学习和模型预测控制的方法,可在保证约束的同时提升控制性能,对腿式机器人控制有潜在应用。


← All digests

Comments