Reproduce “Steer2Grasp: Inference-Time Embodiment-Aware Steering for Diverse Physically Feasible Grasp Diffusion” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Vignesh Vembar, Ayush Kaura, A Padmaprabhan, Siddharth Sinha, Kailash Nagarajan, Keshab Patra, Md Faizal Karim, K Madhava Krishna
Steer2Grasp steers a population of grasp candidates toward robot-and-scene feasibility while keeping its pretrained generator frozen. It scores predictions for reachability and collision avoidance, then reweights and resamples candidates rather than locally correcting grasps. In MuJoCo, CS@all rises from 15.60% to 70.42% for single-arm and 4.88% to 50.77% for dual-arm grasps; this measures reachability and collision avoidance, not stability or task success.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.