Robotics Briefing /
Which obstacles should a robot move?
A cart blocks the way. If it holds an empty bin, a robot might push it aside. If it carries something that could topple, the same push could make a mess. Three studies from this week put navigation to a physical test, from choosing what to touch to following a light that keeps moving.
Your quick takeaways
- A mobile robot reached its goal safely in 10 of 15 scenes when it first judged which objects it could move. Asking its planner to decide everything at once worked in six.
- A rover’s depth camera made open ground look blocked beyond about three meters. The team caught the error indoors before running two outdoor missions.
- A drone followed a cart by reading its blinking lights. Its flight error grew from about three centimeters during hovering to about fifteen when the cart drove in a circle.
Papers
A suitcase with books on top changes the route
Kenneth Llontop and colleagues · November 5 arXiv posting · Accepted at CoRL 2026
In one MessyNav trial, a suitcase topped with a stack of books sat near a cart carrying one book. Pushing the suitcase risked knocking the stack over. The robot moved the cart. Its advantage was the order in which it made two decisions. First a model looked at the objects and judged which to leave alone, pick up or push. Then a planner used those judgments to choose a route.
The team compared four methods in a bedroom and a conference room. Each got one attempt in each of 15 scenes. MessyNav reached the goal without an unsafe interaction in 10. The same planner without the separate object judgment succeeded in six. A planner that treated obstacles according to their height succeeded twice.
There was a cost to relying on visual clues. In a separate test where people judged plans from images, the extra step helped overall but hurt at night. And once the real robot began a push, its routine could not adjust the force in response to what it felt. A sensible choice of object still leaves the physical handling to solve.
Why a rover saw a wall where it could have driven
MSU Rover Team and collaborators · November 5 arXiv posting · Accepted at RiTA 2026
The rover’s assignment was to follow arrows, stop at an orange cone and return to its start. In an indoor trial, its map treated space beyond its depth camera’s useful range as occupied. A route planner working from that map had a good reason to stop, even when the way ahead was clear. Filtering the depth readings corrected the map.
The paper, posted this week, documents tests from November and December 2025. Across one indoor and two outdoor missions, the team recorded returns to the start without collisions or driving commands from an operator. The outdoor trials ran over grass and fallen leaves in light rain.
How close did the rover come to its starting mark? Its software accepted a return within 40 centimeters of the stored location. The team had no independent measurement of the final gap on the ground. The three runs show the cameras, mapping software and steering working together, while leaving the physical return accuracy an open question.
Blinking lights give a drone something to follow
Junkai Niu and colleagues · November 5 preprint · Physical flight tests
The lights on the team’s cart blink at the same rate, but each stays on for a different fraction of a blink. A camera aboard the drone records changes in brightness as they happen. That lets the drone tell the lights apart and estimate where the cart is and which way it faces, even as the pattern moves across the camera’s view.
Flight measurements expose the difference between seeing the target and keeping up with it. With the cart stationary, the drone’s trajectory error was 2.6 centimeters. With the cart driving straight it was 13.3 centimeters, rising to 14.9 on a circle under changing light. The authors attribute much of the gap to the time the flight controls took to respond.
This approach requires lights mounted on the target in a known arrangement. It offers a way for two equipped machines to coordinate. The authors describe a code and dataset release as future work.
News
AIR puts the heavy parts of a factory robot in its torso
Embodied AI · November 6 unveiling in Munich
Embodied AI introduced AIR, a wheeled factory robot whose motors and gearboxes sit in its torso. Tendons transmit motion to the arms. The company says the lighter arms can yield on contact. It also describes how a new task begins: a human operates the robot, the system records demonstrations, and an operator stays available once the robot starts doing the work itself.
The company cites 98% reliability on a cable-handling test with an electronics manufacturer. That result came from its earlier two-arm industrial cell, which uses the same learning software. AIR itself is undergoing manufacturer testing, and the company says CE marking is planned for 2027. Two figures that would help assess the new machine are how often it needs an operator and how long each intervention takes.
Read this first
Open the annotated conference-room scene on the MessyNav project page. Before looking at the labels, decide which object you would move. Then compare your reasoning with the robot’s: the contents of the cart and what rests on the suitcase matter more than either object’s name.
Read the paper