ArXiv robotics surge

Robots still struggle with the last inch of a task: lining up a plug, coordinating two arms, or feeling a surface through a soft sensor. New arXiv papers posted on April 22 and April 23 target those exact bottlenecks. (arxiv.org 1) (arxiv.org 2) One paper, “Learning Hybrid-Control Policies for High-Precision In-Contact Manipulation Under Uncertainty,” was submitted on April 22 by researchers at Oregon State University. It trains a Franka FR3 robot in IsaacSim and IsaacLab, with simulation physics and low-level control running at 120 hertz for connector-style insertion tasks where force mistakes can damage parts. (arxiv.org) That paper tackles a common factory problem: vision can tell a robot where an object is, but not always how hard it is pushing once contact starts. The authors say pose-only policies often need carefully tuned low-level controllers in delicate insertion tasks, and their hybrid setup adds more explicit force handling during contact. (arxiv.org) A second paper, “Bimanual Robot Manipulation via Multi-Agent In-Context Learning,” was submitted on April 23. It uses large language models as action planners for two-arm tasks, splitting the job across multiple agents because standard in-context learning can run out of room when both arms must coordinate many joints at once. (arxiv.org) In plain terms, in-context learning means giving a model examples in the prompt instead of retraining it. The authors say that approach can preserve generalization for robot control, but bimanual manipulation overloads a single model’s context window, so they distribute decisions across agents to manage the action space. (arxiv.org) A third paper, “ETac: A Lightweight and Efficient Tactile Simulation Framework for Visuotactile Sensors,” was also posted on April 23. It focuses on touch simulation, where soft rubber-like sensor skins deform on contact, and says existing simulators often force a tradeoff between physical realism and the speed needed for large-scale policy training. (arxiv.org) ETac models elastomer deformation with a data-driven propagation method rather than a fully heavy soft-body simulation for every interaction. The authors say that keeps simulation quality high enough to capture global deformation while boosting efficiency for training manipulation policies that depend on touch. (arxiv.org) A fourth paper drawing attention this week, “DeVI: Physics-based Dexterous Human-Object Interaction via Synthetic Video Imitation,” was submitted on April 22. It uses video diffusion models to generate synthetic human-object interaction videos, then feeds that motion into a physics environment as a planner for dexterous manipulation. (arxiv.org) That approach aims at a data shortage that has slowed hand manipulation research for years: high-quality recordings of complex finger movements are hard to capture at scale. The DeVI authors say synthetic videos can cover a wider range of scenarios and object categories than motion-capture pipelines alone. (arxiv.org) The timing is notable because the next IEEE International Conference on Robotics and Automation, the field’s flagship meeting, is scheduled for June 1 to June 5, 2026, in Vienna. arXiv is not peer review, but it is where robotics groups often post results first, and the archive now hosts nearly 2.4 million scholarly articles across fields. (ieee-ras.org) (arxiv.org) Taken together, the week’s papers center on the same unsolved problem: getting robots to handle contact, coordination, and touch with less brittleness. The next test is whether these methods hold up outside simulation and preprints as conference season gets underway in June. (arxiv.org 1) (arxiv.org 2)