This Spiral Robot Taught Itself to Grab and Throw Things
A squishy robot that learns by trial and error could soon sort your packages.
Researchers built a spiral soft robot that can grasp objects and pick-and-throw them starting from an ungrasped state — without prescribing contact forces, contact locations, or body configurations. Grasping is measured by tip angular sweep and body-object enclosure, and throwing adds release-direction alignment and a minimum release speed. Because the actuation-to-outcome mapping is nonsmooth, the team used derivative-free CMA-ES in an optimize-learn-refine framework: CMA-ES generates solutions for sampled conditions, a task-conditioned predictor learns warm starts, and CMA-ES refines them for unseen conditions. In simulation, the method achieved 492/500 successful grasps (98.4%) and throwing success rates of 98%, 97%, and 94% across three directional throwing trials. Learned initialization increased grasping success from 78.6% to 98.4% and reduced the median CMA-ES rollout count from 1184 to 816. In hardware, grasping succeeded 100% across 50 executions and pick-and-throw 100% across 30 executions, with 10 repetitions per direction.
- A soft spiral robot uses its whole bendy body — not a claw — to wrap around objects and toss them, no pre-set grip required.
- Adding a learning shortcut boosted grasping success from 78.6% to 98.4% while cutting practice attempts from 1,184 to 816.
- In physical tests it succeeded 50 out of 50 grabs and 30 out of 30 throws, but only for the specific objects tested in the lab.
Why It Matters
Could lead to robots that safely sort fragile packages, produce, and recycling — jobs that are hard to automate today.