Structures-Computer Interaction Lab of UCLA

We split the task in two: Approach: a learned policy guides the arm through the plants to a spot just in front of the target. Interact: a recorded close-range motion finishes the job. Up close, depth sensing gets unreliable, so replay is steadier. #Robotics

Our main idea: before the robot decides how to move, label each camera pixel as one of three things: Target: reach it Obstacle: avoid it Background: ignore it Add depth, and the robot works from a simple 4-layer map instead of a busy photo. #ComputerVision #RobotLearning

Many robot demos happen on clean tables. A greenhouse is different: dense leaves, thin stems, changing light, and plants that keep growing. A robot that simply copies human demos tends to head straight for the fruit and brush through the plants around it. #Robotics #AgTech

Stairs are no problem for our snap-actuated robot: it hops its way up and back down. From our recent Science Advances paper on helical snap-through actuators. Video: https://lnkd.in/ggcsU-Ub Paper: https://lnkd.in/ehvSimVU

What should a robot reach, avoid, or ignore in a cluttered scene? Our #CoRL2026 paper ObstaDiff: an obstacle-aware diffusion policy with 75.4% success and 8.2% obstacle collisions over 61 real-robot trials. Visit our poster in Austin and let's chat! 📄 https://lnkd.in/gdgn_YCg

Snapping elastic limbs vs. rigid legs: across six terrains, our snap-actuated robot averaged 2.46 body lengths/s, compared with 0.79 for its rigid-legged counterpart. From our recent Science Advances paper. Video: https://lnkd.in/gMWBRAwK Paper: https://lnkd.in/ehvSimVU

Our recent work (Science Advances) turns snapping helices into robot limbs. The result is an untethered, frog-like robot, 11 cm long and 98 g, that can hop, flip, and swim using only a small low-power motor. Video: https://lnkd.in/g2ZsVxsj Paper: https://lnkd.in/ehvSimVU

A motor slowly winds up the helical limb, the limb snaps, and the cycle repeats. This repeatability test shows the snap stays consistent over many cycles, turning slow input into fast motion. Video: https://lnkd.in/gbM_MTe4 Paper: https://lnkd.in/ehvSimVU

With the right geometry, twisting an elastic helix stores energy that is released in one sudden snap. The video shows the nonlinear snap-through response of the elastic limb. Video: https://lnkd.in/eUA2TdvJ Paper: https://lnkd.in/ehvSimVU

Our recent work in Science Advances shows that geometry decides whether an elastic helix buckles gradually or snaps suddenly. Here, a helical filament buckles as its ends are bent and twisted. Video: https://lnkd.in/ewJHAWgX