Research
FORGE: Force-Guided Exploration for Robust Contact-Rich Manipulation under Uncertainty
- Input
- Robot proprioception plus force/torque sensing and (uncertain) pose estimates, for contact-rich assembly tasks (peg-in-hole insertion, gear meshing, nut threading) under real-world pose uncertainty that sim-only training doesn't have to deal with.
- Supervision Required
- Reinforcement learning in simulation, with two mechanisms doing the real work instead of hand-labeling: a learned force-threshold that decides when contact force means "back off" versus "push through," and dynamics randomization so the sim-trained policy survives contact with the real world.
- Representation Learned
- A force-aware manipulation policy that fuses proprioception and force/torque feedback with a learned force threshold, rather than treating force purely as a penalty term, letting it distinguish expected contact during insertion from contact it should react to.
- How It Connects to Robot Actions
- Direct and this is the connection I'm building toward, not one I've fully closed yet. My peg-insertion project trains against Isaac Lab's `Isaac-Forge-PegInsert-Direct-v0` environment (FORGE's own benchmark task), but my ablation only tests force *observation* vs. force *penalty*. I haven't implemented FORGE's actual force-threshold mechanism or dynamics randomization scheme yet.
- Real Hardware Test
- FORGE itself validates sim-to-real transfer on a real robot arm across multiple assembly tasks, including a multi-stage planetary gearbox assembly. My own implementation has stayed sim-only in Isaac Lab so far.
- Limitations / Failure Modes
- Force-threshold tuning and dynamics randomization both add real engineering surface area: get the randomization ranges wrong and sim-to- real transfer fails quietly. My own results underscore the underlying fragility even before adding FORGE's specific mechanisms: a plain force penalty without observation made performance *worse*, not better.
- Experiment I Would Add
- Implement FORGE's actual force-threshold and dynamics-randomization scheme on top of my existing peg-insertion setup, then attempt real sim-to-real transfer on a physical arm instead of stopping at the sim-only ablation.