Onasafari

Whole-arm grasping, revisited: my 2014 MIT master's thesis (Learning a Strategy for Whole-Arm Grasping), re-run with 2026 tools. Same spirit — a 3-DOF arm learning enveloping grasps, force closure as the goal — harder task: language-conditioned targets in clutter that I singulate before grasping. Tell the arm what to grasp in words: a VLM grounds the instruction, and SAM 3.1 (Meta Model API) segments the target. An antipodal sampler proposes grasps on the mask, and MuJoCo executes. My thesis ran Box2D with SARSA; here PPO from scratch, BC, and diffusion policies join scripted and VLM planners. I benchmark all seven planner families on identical seeds across five clutter scenes, scored on force closure, singulation, and displacement. Headlines: PPO wins 40/50 by learning restraint; BC ties its teacher 40–40 at 2× data. Docker-reproducible, 115+ commits Oct 27–Nov 4.