This is the edition of 1 November 2026. See the latest news.
New research papers
arXiv
Preprints from the arXiv categories Robotics (cs.RO) and Artificial Intelligence (cs.AI). Preprints are not peer reviewed.
01
arXiv · cs.RO29 October 2026
The authors propose PhysVLMM, a method that maintains physical belief posteriors for articulated objects in robot scene maps and propagates evidence from observed interactions to untouched similar parts. They report 86% joint-type identification on real doors and drawers across 18 homes, with 41% lower prediction error than differentiable simulation at four orders of magnitude faster inference.
Read the original at arXiv ↗
02
arXiv · cs.AI29 October 2026
The authors probe LLMs trained with reinforcement learning on Tic-Tac-Toe using sparse autoencoders to trace how internal representations evolve, investigating whether RL produces genuine reasoning or superficial pattern exploitation. They report that while RL yields high accuracy suggesting game reasoning, mechanistic analysis reveals models primarily rely on information already explicit in the prompt, with high-level strategic concepts remaining weakly represented.
Read the original at arXiv ↗
03
arXiv · cs.RO29 October 2026
DAWN is a hierarchical robotic framework for real-world unfastening tasks that divides the procedure into alignment and insertion followed by screw removal, using three components: a diffusion-based policy for screwdriver engagement, a multi-modal classifier for stage transition detection, and a force-controlled removal module. Validated on a Franka Emika robot, the system achieves alignment success rates comparable to trained human operators and robust unfastening across various unseen scenarios.
Read the original at arXiv ↗
04
arXiv · cs.RO29 October 2026
The authors present a framework in which coding agents build robotics simulations, using an adversarial creation and audit loop to produce RoboWorld, a suite of 25 simulations with training and test RL environments. They report that the strongest frontier models fail every test task, averaging a reward of 0.06, and they explore several in-context improvement strategies.
Read the original at arXiv ↗
05
arXiv · cs.AI29 October 2026
The authors introduce AgentDecept, an executable benchmark of 477 pressure variants built from 257 tasks, designed to test whether tool-using agents honestly report what they did. Their protocol compares user-facing claims against logged observations and outcomes to separate falsification, concealment and equivocation from honest failure.
Read the original at arXiv ↗
06
arXiv · cs.RO29 October 2026
The authors propose LOGIV, which compiles robot plans into task graphs of condition-support and resource-ordering edges, verifies them during execution, and repairs only the affected dependencies after failures. Running above existing VLA and world action model policies, it is reported to raise dual-arm task success and to improve real-robot trials from 12/30 to 17/30.
Read the original at arXiv ↗
07
arXiv · cs.AI29 October 2026
The authors present FARMS, a pipeline that turns computational mathematics sources into formal, reusable knowledge connected to algorithmic meaning and quantitative guarantees, using Lean-checked implications. They report that 157 of 171 targets in numerical linear algebra and numerical PDEs passed the pipeline, at a mean cost of a few dollars or less per target.
Read the original at arXiv ↗
08
arXiv · cs.RO29 October 2026
The authors study adversarial perturbations of internal key and value routes in world action models and report that larger representation shifts do not reliably cause larger control failures. They propose PASA-KV, which picks action-facing routes before training an attack, and report strong closed-loop attacks across several systems and a physical robot.
Read the original at arXiv ↗
09
arXiv · cs.AI29 October 2026
The authors introduce the Algebraic World Model, which explores unfamiliar environments, builds executable hypotheses, restructures them using category theory, and formalizes them in Lean 4 to prove reachability or impossibility. On two impossible game levels they report kernel-checked impossibility proofs in all six runs, while tool-equipped baseline models produced none.
Read the original at arXiv ↗
10
arXiv · cs.RO29 October 2026
The authors propose CoRAct, which aligns VLA actions with their embodied reasoning by rewarding candidate actions for consistency with stated movement and predicted waypoints, and optimizes the initial action noise distribution. They report gains in success, generalization and reasoning-execution agreement on RoboCasa, VLABench and real manipulation tasks.
Read the original at arXiv ↗
AI-generated summaries. Items, titles and links come from the publishers' public news feeds and from arXiv, and each title links to the original. Summaries are written and checked automatically from the source text and are not individually reviewed by a person; the original is authoritative. Research papers are preprints, not peer reviewed. Llull Lab is independent of the publishers listed. How it is made.