01dynamic manipulation
RAI Institute researchers built light, backdrivable robot arms for fast physical interaction. AthenaZero throws balls at more than 30 metres per second and catches or bats faster pitches in controlled tests. The hardware makes a lively game of catch possible, though holding heavy loads still brings thermal limits.
Why it mattersReducing moving mass helps a robot react to fast physical interactions rather than only execute slow preplanned motions.
Keep in mind: These are baseball-inspired controlled tasks. Sustained static loads can overheat motors, and the demonstrations do not prove general-purpose household ability.
Read the source · Andrew S. Morgan and colleagues, RAI Institute ↗02flapping wing robotics
Researchers built flapping robot wings that borrow an idea from insect muscles: motion helps generate the next beat. The wings adapt to mechanical changes and react to collisions through their own actuation dynamics. Laboratory flights through hanging obstacles were more stable than the comparison system, while overall flight still relies on a controller.
Why it mattersCarefully designed mechanics could let flying robots react to contact before a central controller has time to intervene.
Keep in mind: These are controlled laboratory flight and collision tests. The wings respond through their actuation dynamics, but the robot still needs a flight-control system.
Read the source · Rundong Yang and colleagues, UC San Diego and collaborators ↗03robot learning
AnyViewDex researchers trained a robot hand with geometric information in simulation, then deployed it using a single color camera. On an xArm7 with a LEAP Hand, they report 76.7% grasping success across eight unfamiliar objects and six camera viewpoints. It is a controlled grasping result, with occlusion and clutter still difficult.
Why it mattersA robot that tolerates camera movement could be less dependent on a carefully calibrated laboratory setup.
Keep in mind: The hardware result covers 480 main trials in a specific grasping setup. Total occlusion, small objects and clutter remain limitations.
Read the source · Soham Patil and colleagues, IIIT Hyderabad and collaborators ↗04navigation learning
Northeastern researchers collected 37.2 kilometres of routes using a rollator walker and smartphone. Because the walker cannot climb stairs or uncut curbs, its paths reflect wheel-friendly choices. Models trained with that data transferred to a powered wheelchair in selected scenarios. The collection tools and full dataset are still awaiting release.
Why it mattersA cheap physical proxy can collect routes that respect wheel constraints without needing the target robot for every demonstration.
Keep in mind: Transfer demonstrations cover curbs, stairs and curb cuts. Code, trained weights and the privacy-processed dataset are not yet public.
Read the source · Sarvesh Prajapati and colleagues, Northeastern University ↗05robot force control
A Barcelona team used paired robot arms to record how firmly a person intended to press during a demonstration. A trained robot then pushed harder when asked for a firm wipe. That is a step toward teaching touch through examples, though the evidence comes from one wiping task and one robot platform.
Why it mattersUseful physical instructions need to express how hard to push, as well as where to move.
Keep in mind: Evidence covers one wiping task on one robot platform with a small rollout budget. The causal contribution of the force input remains unresolved.
Read the source · Harsha Guda, Adrià Colomé and Carme Torras, CSIC-UPC ↗06world models
WorldContact researchers trained a model to predict contact with flexible shopping bags and generate extra robot-training data. A real arm's single-attempt lifting success rose from 65% to 95% in their comparison. The wider task set remains simulated, and the reported generation speed excludes rendering and saving files.
Why it mattersA learned model of contact can generate more training experience for flexible objects that are difficult to simulate quickly.
Keep in mind: The tenfold speed comparison excludes rendering and disk I/O. Real-robot validation concerns a narrow bag-lifting task, not all 16 simulated tasks.
Read the source · Caoliwen Wang and colleagues, UBC, Westlake and Style3D ↗