Specialty
A preprint reports that MILD increased a robot’s foot clearance on softer surfaces and its stride length during an unannounced shift from high to low stiffness. The robot completed 10 forward-and-backward cycles at 1.2 m/s on each of seven terrain types without a reported failure, though independent trial counts and variability were not fully reported.
IEEE Robotics and Automation Letters (2025)3 min read
A robotics preprint tests PVRA, a supervised RGB-D framework for estimating target and assembly poses during progressive robotic assembly. In synthetic Nema17 scenes, PVRA’s primary pose score was 0.893, compared with 0.989 for FoundationPose using ground-truth masks.
Scientific publication3 min read
A preprint reports a quadruped robot jumping through narrow gates, reaching up to 2.5 metres per second in the described traversal. The evaluation also covered shifted gate positions and additional terrain and dynamic-task scenarios, but it did not report a complete overall success rate or numerical details for the comparisons.
Advanced Robotics Research (2025)4 min read
An arXiv preprint tests EAFG, a robot-planning approach that gathers visual evidence before task planning. It reports higher full-recipe completion for underspecified instructions and higher halt rates when a required object was missing, alongside mixed results in explicit tasks.
Scientific publication3 min read
A new arXiv preprint maps the shift in end-to-end autonomous driving from copying low-level controls toward systems built around more structured planning. Its central warning is that prediction accuracy or impressive language reasoning alone cannot establish safe driving.
Scientific publication5 min read
An arXiv preprint describes a suspension-preconditioning method for off-road vehicle jumps. In selected BeamNG.tech simulations, the combined controller recorded higher success counts than comparison policies, while other tested speeds and ramp profiles exposed important reversals.
Scientific publication3 min read
A 2026 arXiv preprint compares 41 latent-action design choices for robot learning. LAPO had the highest reported overall benchmark mean, while several design choices were associated with higher scores. A nonrandomized real-world comparison does not establish a causal advantage.
Scientific publication4 min read
A preprint describes GOAG, a gripper-centric grasp planner designed to generalize across unseen object shapes. It reports an 86.93% average simulated success rate on MultiDex, a 53.97% average across five benchmarks and a successful demonstration on 11 YCB objects, while leaving questions about matched comparisons, uncertainty and wider physical testing.
2026 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)5 min read
A robotics preprint describes a pipeline that turns smartphone video into a simulated door twin, generates demonstrations in simulation and deploys a depth-based policy on a wheel-legged mobile manipulator.
Scientific publication4 min read