Viewpoint
MAM system combines short-term visual and long-term textual memory for robots.
MAM system combines short-term visual and long-term textual memory for robots.
Marcel presents MAM (multi-scale embodied memory), a system that decomposes memory into short-term visual memory for low-level dexterous control and long-term textual memory for high-level policy planning, enabling robots to track task progress and adapt to mistakes during long-horizon tasks.
- Speaker
- Marcel Torne Villasevil
- Source timestamp
- 7:59
More from this interview
- Four technical walls block scaling of robotics according to Francois
- Simulation-trained policy achieves zero-shot real-world tool manipulation
- R&B Encore uses self-supervised chain-of-thought reasoning for VLAs.
- Robotics application companies should own problems and build operational moats.
- World action models are computationally expensive, a barrier to deployment
- Distillation and separate transformers speed up world action model inference