Robotics Briefing /
A robot arm gets a second try
A robot packing cans into a bin bumps something and stops. Does that end the attempt? In an Apple study, the answer changed how often the arm finished the job. Two other teams looked at decisions that come earlier: where to take hold of a chair, and how a robot without a hand can move a box at all.
Your quick takeaways
- Apple’s robot controller finished 104 of 108 physical trials when it could request recovery after a stop. It finished 94 when the same kind of stop ended the attempt.
- In simulation, a robot that chose its grip to suit the destination moved eight kinds of objects into place about 90% of the time. Picking a grip at random yielded 74%.
- A sidewinding snake robot pushed a box to within 20 centimeters of its goal in all 20 lab trials. External cameras tracked both the robot and the box.
Papers
What happens after the arm stops?
Chen Chen and colleagues at Apple · November 6 preprint · Physical robot tests
AMP1 lets a general-purpose AI model inspect camera images, write short programs to move an arm and its gripper, and check what happened before choosing another movement. The researchers supplied ways to see and move, leaving the model to work out how to grasp or place an object. It received no demonstrations of the test tasks.
The clearest comparison comes from packing cans. When a bump triggered the arm’s protective stop, the model could ask the controller for its status and request recovery. It could then try a different movement if the controller allowed it to resume. With that option, the arm packed the cans in all nine trials. When a controller error ended the trial, it finished three of nine.
Across twelve tasks on three physical robot setups, recovery raised AMP1’s total from 94 to 104 successes in 108 trials. The comparison systems also had to stop on a controller error. Speed is a separate question: getting the lemons and an orange onto a plate took AMP1 an average of 190 seconds across its nine successes. Apple has released the software used to run its benchmark evaluations.
Which part of the chair should the robot hold?
Xun Tu, Minghao Zou and Karthik Desingh · November 6 preprint · Simulation
Either armrest of a chair might give a robot a secure hold. Moving the chair to the left and turning it to face a desk could still be easier from one side than the other. Grasp-TAGS learns to rank those starting positions by what happens at the end of the move. During practice, the robot tries different grips and destinations. A small model then uses that experience to pick a grip before the robot makes contact.
The researchers tested three chairs, two carts and three buckets in simulation. The goal specified both a position and a final orientation. Across the eight objects, choosing the grip for the goal yielded 89.84% success. The best policies trained for a single grip averaged 79.87%. Using the same multi-grip controller but choosing its starting grip at random yielded 74.07%.
The researchers supplied a shortlist of feasible grips for each object. Finding those grips from camera images in an unfamiliar room would add another problem, and testing the method on a physical robot remains future work.
A snake moves a box with the same body it walks on
Adarsh Salagame and colleagues · November 6 arXiv posting · Presented at IROS 2026
COBRA has eleven powered joints and no need for a separate gripper in this task. As the robot sidewinds, one section of its body can press against a box while other sections push against the floor. To steer it, the researchers planned where the body as a whole should go. This let the controller guide the push without having to plan every oscillation of every joint.
The physical robot brought the box within 20 centimeters of its goal in all 20 trials across different starting positions and destinations. One longer trip took the box through two intermediate waypoints. The lab’s motion-capture system told the controller where the box and robot were throughout the trip.
The team separately tested a mathematical model of the forces on the box in simulation. The physical trials used the simpler steering system. Tracking the box with the robot’s own sensors and adapting the gait to uneven ground are still open steps.
News
Robot learning researchers meet in Austin
Conference on Robot Learning · November 9–12
The main Conference on Robot Learning is scheduled for November 9–11, followed by workshops on November 12. Two sessions connect directly to the questions in today’s papers: one on using tools with both hands, and another on keeping control of an object while the robot itself moves. The organizers have published the workshop links, including the questions each group plans to tackle.
Read this first
Try the banana-and-bowl walkthrough on Apple’s project page. A banana curves, so the average position of the points seen by a depth camera can fall in empty space. Follow how the model chooses a point on the fruit near that average. It’s a small decision that makes the difference between closing the gripper on the banana and closing it on air.
Read the paper