This Soft Robot Has No Gripper. Its Body Does the Grabbing.

A soft, rubbery robot with no fingers used its entire body to grip an object, lift it, and release it in a chosen direction. Across 50 real grasps and 30 directed throws, it did not miss.

The robot is the work of researchers publishing on arXiv in late September 2026, and the result is narrow on purpose. The body is a spiral-shaped piece of soft material, actuated by tendons. Instead of a hand, the researchers taught the robot to bend, wrap, and let go, so the body itself becomes the gripper. The paper is Optimize, Learn, Refine: Whole-Body Grasping and Pick-and-Throw with a Spiral Soft Robot.

What is harder is making one reliable. A soft body is a liability for precision: it bends, sags, and deforms in ways a rigid arm does not. A traditional control loop, where the robot senses its own position and adjusts, becomes brittle on compliant material. The authors' answer is to drop the feedback at the task level and lean on a search loop instead.

The system has three pieces, in plain language. First, the robot sweeps its tip across the object and watches how close the body comes to it, so the system measures a real "how wrapped am I" signal rather than just a position. Second, when the job is a throw, the system adds two extra objectives: which way the object should leave the hand, and how fast it should be moving at release. Third, each new task gets a starting guess that is learned from prior attempts, so the search does not begin from scratch.

The third piece is the warm-start loop. The search method, a population-based optimizer called CMA-ES, normally needs many trials. With a learned warm start, the median number of rollouts dropped from 1,184 to 816, and the simulated grasp success climbed from 78.6% to 98.4% on 492 of 500 simulated grasps. The authors describe the loop as "optimize, learn, refine": generate solutions, learn which starting strategies transfer, then refine on the new task.

In the real robot, the authors report 50 of 50 hardware grasps and 30 of 30 directed throws, with 10 repetitions per throwing direction. The 78.6%-to-98.4% jump is the simulation's contribution to the warm-start story.

The studied tasks are explicitly planar: motions happen in a 2D plane, not in full 3D. Tendon commands run open loop at the task level, with no updates from object state or contact feedback while the motion is in progress. Success is measured against defined tasks and target zones, not against clutter, varied lighting, or unstructured environments. "Open loop" here means the robot does not react mid-motion to what the object is doing; it commits to a plan and lets the soft body absorb the variation.

Compliant bodies fit places rigid arms cannot, wrap around irregular shapes, and tolerate soft or fragile items a parallel-jaw gripper would crush. The paper's version of that promise is bounded: one robot, one lab, simple controlled objects, a small set of throwing directions. There is no claim of deployment, no warehouse trial, and no timeline.

The warm-start loop carries the result, and the paper itself stops at planar tasks and open-loop execution. The next read is the one that adds a vertical axis, a moving object, or a less controlled target zone, which are the parts the current result is built to leave for later. The full preprint is available in the HTML version.