· W45
Week 45: DLO Keypoint RL, FlowBench, Clinician-AI CAMs
A robot can practice a task inside a computer. The fake world lets it try many times without harm. Then it must repeat the same task on a real machine. Small gaps between the two settings cause trouble. The robot has to adjust what it learned.
DLO Keypoint RL
- did
- The policy moved a one-end-grasped deformable linear object to target keypoints on a planar manipulator.
- transfer
- It transferred zero-shot with markerless RGB-D keypoint estimation, succeeding in 9/10, 10/10, and 8/10 physical trials.
- sim got wrong
- Simulation lacks real sensing noise and contact mismatch, so training added noise, delay, domain randomization, and disturbances.
- unmeasured
- The paper does not report real-world execution time or error magnitudes.
arXiv:2611.07261 · 2026-11-05
FlowBench
- did
- FlowBench built 428K training examples for counterfactual spatial imagination and tested vision-language models and world models.
- transfer
- They fine-tuned Qwen3-VL-8B on FlowBench-Sim-Train and tested on real images, gaining 16.9 points.
- sim got wrong
- Simulation gives exact outcome frames and optical flow, which real images cannot, so they train on simulation.
- unmeasured
- The paper reports no real-robot trial count or hardware success rate.
arXiv:2611.03937 · 2026-11-02
Clinician-AI CAMs
- did
- The in-vitro study tested six clinician-AI robotic navigation modes in a patient-specific vascular phantom with 11 experts and 2 students.
- transfer
- Autonomous navigation used zero-shot transfer from simulation to the phantom, dropping from 81% success to 23%.
- sim got wrong
- Simulation overestimated full autonomy, so shared control modes raised phantom success to 92-97%.
- unmeasured
- The paper does not report per-mode trial counts or confidence intervals.
arXiv:2611.03446 · 2026-11-02