· W44
Week 44: catching, a slope, and a surgical handoff
Each paper fixes a different part of the problem. The catching hand trusts nothing about contact. The rover distrusts the ground. The surgical robot invents the handoffs it needs.
All three work around the simulation instead of measuring it. That is useful engineering, and it leaves the same hole in every paper: no score for the same task in simulation and on hardware.
Catch It All!
- did
- Caught nine unseen objects with a Unitree G1 and one head camera: soft toys, foil bags, hard objects that bounce.
- transfer
- Zero-shot. No training on the real robot.
- sim got wrong
- Contact forces. The reward was shaped so the hand survives contact it cannot predict, and wrist torque was capped at 5 Nm.
- unmeasured
- How much worse it did on real objects than in simulation.
arXiv:2610.39044 · 2026-10-28
TERRA
- did
- Learned to cross a loose slope with a rover called ERNEST, after recovering its walking gait in three to five real trials.
- transfer
- Online adaptation. About 90% slope success within five simulated trials.
- sim got wrong
- Slip and sinking. The rover collects real data where its model is most wrong instead of simulating the ground better.
- unmeasured
- The same slope scored in simulation and on hardware.
arXiv:2610.38657 · 2026-10-28
SimRetract
- did
- Took a retractor from a surgeon's hand in open surgery and held it, working from partly hidden instruments.
- transfer
- Zero-shot. Simulated demonstrations only, no real demos, no fine-tuning.
- sim got wrong
- Human contact. The simulated handoffs include failures, so the robot has practised recovering.
- unmeasured
- What happens when the surgeon moves somewhere unexpected.
arXiv:2610.38234 · 2026-10-28