Astribot’s Lumo-2 uses latent world dynamics to improve long-horizon robot tasks

jiqizhixin · x · 2026-07-27

Astribot’s Lumo-2 introduces a latent world-action model for robotics that reasons about world dynamics before acting.

• The model generates actions in a compact latent space rather than relying on simple memorization.
• It uses multi-stage alignment to coordinate vision, language, and action representations.
• The team reports consistent gains over top vision-language-action and world-action baselines.
• The strongest improvements appear on long-horizon and dexterous manipulation tasks, where temporal reasoning and physical understanding matter most.
• The paper frames the work as a step toward predictive, aligned, and scalable robot learning.

Original post →

More from Embodied

• XPENG's IRON robot demos full-duplex speech with 9-mic array and lip reading — ChrisGPT · 2026-09-23
• Uber riders can now hail Waymo robotaxis on Austin freeways — ATTlKA · 2026-09-23
• M5 Ultra LLM test: 4x faster prompt processing, but double the power draw — DigitalguyCH · 2026-09-23
• Figure CEO: Humanoid Robotics Needs Four Stages and Eventually Hundreds of Billions — adcock_brett · 2026-09-23
• InstinctFlash: open-source runtime runs 8 robot model families in real time on one commercial GPU — chris_j_paxton · 2026-09-23
• M5 Ultra 96GB vs RTX 5090-6000 Pro: which rig for local video generation at $6.5K-$27K — MaxwellHusk · 2026-09-23

Embodied channel →