Long-WAM from NVIDIA/MIT: 19s context lifts real-time robot control to 78.7% at 107ms per chunk
songhan_mit · x · 2026-10-14
Long-WAM, from NVIDIA, MIT, HKU and UCSD, scales the visual history of a causal world-action model while keeping real-time robot control.
Key findings:
- Longer context, better control: on RoboCasa GR-1, extending context from 2.4s to 19.2s raises success from 66.3% to 78.7%; LIBERO-Long goes 94.5% → 99.5%.
- AR pretraining matters: causal prediction pretrained on 10,000 hours of robot/egocentric video (no action labels) scales with context; bidirectional initialization shows no net gain.
- Smoother async control: RoboTwin 2.0 hits 94.2% async vs 94.4% sync, with action jerk 3× lower than Fast-WAM and 15× lower than LingBot-VA.
- Edge-ready runtime: the full predict-then-act model runs at 107.4 ms per action chunk on an RTX 5090, deployed on DGX Spark and Jetson AGX Thor.
- Real-robot demo: 95% (19/20) on dynamic cup stacking with Unitree G1.
More from Embodied
- Physical Intelligence researcher: robotics is in a 'Cambrian Explosion' — svlevine · 2026-10-14
- Robot demo nails 20/20 picks: nuts, gears and connectors in one run — KyleMorgenstein · 2026-10-14
- General Instinct partners with Physical Intelligence to deploy frontier models on real robots — svlevine · 2026-10-14
- Physical Intelligence opens Pi robotics model API beta, already in 150 factories — svlevine · 2026-10-14
- The 'sophisticated device' in spy case is a $14.90 Seeed Studio ESP32 hobby kit — wavefnx · 2026-10-14
- KILVO: robust multimodal odometry for humanoid robots fusing encoders, IMU, LiDAR and vision — rsasaki0109 · 2026-10-14