NUS's Grounded Action Model Tops Robot Manipulation Benchmarks with 3D Grounding
NationalUniversityofSingapore · hf · 2026-09-22
NUS researchers propose Grounded Action Models (GAMs), a robot foundation model paradigm built on 3D grounding. Unlike VLA and world-action models whose pretrained backbones learn metric grounding only implicitly from demonstrations, GAM conditions on language, point, or box prompts, converts selected objects into a shared object-centric representation capturing focused visual features and metric geometry, and mixes it with robot state history in a multi-stream transformer to predict action chunks. It can run autonomously or serve as a low-level controller for high-level planners.
Key results:
- RoboTwin 2.0: 55.3% average success over 50 tasks (vs. 52.0% Spatial Forcing), 47.6% under scene randomization (vs. 30.4% Abot-M0), trained only on clean-scene demonstrations
- LIBERO-PRO: 61% average across 16 perturbations, beating π0.5's 53%, with the largest gains on relocated/re-designated targets
- Real robots: 17/20 successes under visual shift on a bimanual YAM (vs. 4/20 for π0.5); composed with a Molmo2 planner on Franka it reaches 64.7% ID and 49.8% OOD step completion on long-horizon, memory-dependent tasks
More from Embodied
- Designing clothes for humanoid robots: fabric serves the mechanism, not the body — yongqianme · 2026-09-22
- Inside robot data collection: First Contact adds mocap to its setup — Scobleizer · 2026-09-22
- Robot arm grasps cube 71 minutes in: model writes its own kinematics, no training — Scobleizer · 2026-09-22
- Distilling World-Model Features into VLAs: 0.8B Policy Hits 97.9% on LIBERO — Trung Dao · 2026-09-22
- HIRO Industries Unveils Origin: A 17-DOF Dual-Arm Robot for Packing and Kitting Workstations — Scobleizer · 2026-09-22
- Stealth team unveils new robot Origin; Scobleizer bets it would sell at Home Depot — Scobleizer · 2026-09-22