Harvard's GRASP recovers gradients from archived rollouts for dexterous retargeting
hankyang94 · x · 2026-11-01
Harvard's Computational Robotics Group details GRASP, building on their RSS 2026 TSMC work: it combines sampling-based exploration with gradient-guided refinement without requiring differentiable simulator dynamics. The key trick is reusing archived rollouts (including rejected proposals) to fit local cost models and recover gradients — even simple quadratic surrogates help guide MCMC rejuvenation. A Metropolis–Hastings correction with true simulator costs preserves the cost-tilted target distribution, and an MPC formulation extends the optimizer to long demonstrations. GRASP achieves 94.2% success on 1,352 bimanual OakInk2 demonstrations with Sharpa hands, with 2.87 cm mean object position error.
Related event: Harvard Unveils GRASP for Contact-Rich Motion Retargeting(2 posts)→
More from Embodied
- mink 1.4.0 released with new Elastic Constraints feature for robot control — kevin_zakka · 2026-11-01
- Naval predicts 4-passenger drones will eventually end commercial air travel, with a long interregnum ahead — naval · 2026-11-01
- Meta Quest game scans your real room and turns it into a haunted house — Scobleizer · 2026-11-01
- RoboAuto: Robot Program Search With Multi-Level Physical Understanding — DJiafei · 2026-11-01
- Developer dresses robot 'Bob' as a ghost to hand out Halloween candy — chrismatthieu · 2026-11-01
- GRASP authors compare with SuperMimic: extracting more optimization signal from sim experience — hankyang94 · 2026-11-01