AlphaNIST robot training: Newton physics plus sparse reward did the heavy lifting
milesmacklin · x · 2026-10-14
mateoguaman details the improvements from SGS to AlphaNIST: switching from PhysX to Newton for a massive performance boost, better reset distributions, robot controller parameters, observations and network architecture, and a unified multi-task formulation. Surprisingly, these stacked improvements enabled training with a minimal reward function—four general terms with most gradient signal from a sparse task reward, and a single coefficient tuning the entire cost term.
More from Embodied
- Dev ports Wright CAD workflow to ChatGPT Desktop, wiring Codex to SteveCAD for 3D design — burhop · 2026-10-14
- RoboStrategy claims the first AI robot to ring the Nasdaq bell, $BOT draws attention — cixliv · 2026-10-14
- Where Winds Meet to add NVIDIA DLSS 5 with 3D-Guided Neural Rendering in late November — ctnzr · 2026-10-14
- LightNav-0 open-sourced: one VLM-based model navigates four robot types autonomously — rsasaki0109 · 2026-10-14
- RigidFormer accepted at NeurIPS: mesh-free Transformer beats GNN simulators for rigid dynamics — camenduru · 2026-10-14
- Science Corporation implants vision-restoring chip in first Stargardt patient — TechCrunch AI · 2026-10-14