Most robots are trained to avoid touching anything at all. Reach into a cluttered fridge, and a standard robot will freeze rather than risk knocking something over, even if all that’s in its way is an empty water bottle.
A new system built by two USC Viterbi School of Engineering and USC Stevens School of Computing and AI graduate students from the Thomas Lord Department of Computer Science aims to fix that, by teaching robots to make judgment calls and act on them with precision instead of playing it safe by default.
Called IMPACT, it’s the product of over two years of work by 2025 master’s in computer science (artificial intelligence) graduate Karan Owalekar and second-year computer science PhD student Yiyang Ling, and which was recently accepted into the 2026 IEEE International Conference on Robotics and Automation (ICRA). — one of the top venues in AI research.
The work was co-supervised by Daniel Seita, an assistant professor of computer science and the director of the Sensing, Learning, and Understanding for Robotic Manipulation (SLURM) lab at the Thomas Lord Department of Computer Science. Another supervisor was Erdem Biyik, an assistant professor of computer science and electrical and computer engineering at the Thomas Lord Department of Computer Science and the Ming Hsieh Department of Electrical and Computer Engineering.
The idea sounds simple: teach a robot which collisions are harmless and which aren’t. In practice, it’s anything but.
“It also depends on where it is placed in the environment, or how it is going to be interacted with,” Owalekar said. “The glass bottle is risky, but then a plastic bottle next to a glass bottle is equally risky, because if you hit a plastic bottle, it will hit the glass bottle, so the relative positions of everything matter.”
Not all collisions are equal, either. A robot knocking over a crayon is a very different event than one knocking over a pair of scissors — and judging that difference is exactly what Owalekar and fellow USC Viterbi student Yiyang Ling set out to teach a machine to do.
Other systems have tried to give robots this kind of collision reasoning before. What makes IMPACT — Intelligent Motion Planning with Acceptable Contact Trajectories via Vision-Language Models — different is context: the robot reasons on its own without training that a plastic bottle, harmless alone, becomes unsafe to touch the moment it’s sitting next to a glass one.
In engineering terms, IMPACT breaks from “collision-free” motion planning, the long-standing approach that trains robots to avoid all contact rather than reason about which contact is acceptable. Seita, the project supervisor, said the approach challenges some fundamental assumptions in the field.
“It breaks away with standard motion planning and robotics assumptions about what “safe” actions mean,” Seita said. “They were able to effectively implement the research vision while also making detailed adjustments to the method to improve results.”
Owalekar joined Seita’s lab in January 2024, where he first began work on this study. Ling joined him that August, and took on building environments in a physics simulator — an unusually complicated one.
“For most of my prior work, people tried to make the environment as simple as possible — maybe only one object, or at most three objects,” Ling said. “But we put at least five to six objects in one environment. For a physics simulator, this is very difficult because we needed to have a lot of parameters to make it as realistic as possible.”
That simulator was only the first hurdle. Ling and Owalekar then had to carry these complex simulations into the real world, where far more variables complicate the robot’s reasoning — and where Ling’s example from earlier plays out in practice: a robot that can only plan collision-free paths can’t retrieve something buried behind other objects. IMPACT lets it move the obstacle aside first, then grab what it actually needs.
The high-level reasoning IMPACT delivers could bring a better robot to your home.
“Imagine that you want a robot at your home and you want it to grab something for you,” Ling said. “If a robot can only plan a collision-free trajectory, it cannot reach the objects deep in the cupboard… If there’s obstacles in front of the target object, the robot would first pick up the obstacle, place it somewhere else, and then grasp the target object.”
From a research perspective, IMPACT is a motion planning system that puts context first, allowing robots to explore without needing to avoid everything it approaches. To explain IMPACT’s, well, impact, Owalekar offered a second, lighter example: a robot dog. Modern robot dogs, like a typical motion planning system, are often coded to treat everything as an obstacle, even tall grass or tree branches. With a reasoning system like IMPACT, a robot dog navigating a disaster could spot the difference between walking through safe debris and carefully steering around rubble that might collapse. This same idea translates to any environment where collision is unavoidable, like the aforementioned cluttered shelf example.
Owalekar and Ling’s model gives robots the enhanced reasoning power to explore, realizing that what a collision-free model might call an obstacle might in fact be an innocent tree branch.
“If there is a robot who is doing some task and it did something wrong, you can just look at IMPACT and see, ‘oh, this was the reason why that thing failed, and if this is how we can fix it,’” Owalekar said. “It opens up a whole lot of other possibilities… Every decision IMPACT makes can be traced and explained, not only when something fails but for every choice the robot makes along the way.”
Published on July 21st, 2026
Last updated on July 21st, 2026