AD-Memo: language-based memory helps driving agents reason over past observations
drmapavone · x · 2026-11-06
A new paper, 'Vision-Language-Action Autonomous Driving Agent with Language-based Memory,' introduces AD-Memo, a driving agent that stores decision-relevant information about surrounding objects as language-based memory and reuses it in later reasoning.
- Addresses decision-making under partial observability: e.g., remembering that a left car arrived at the crossing 1.6s ago lets the agent correctly decide to yield
- Also supports scene-understanding VQA and plug-and-play memory notes
- Builds on prior work on planning-aligned token compression, which preserves history selectively based on driving intent
The work highlights memory as an emerging research direction for Physical AI agents.
More from Embodied
- Infinite Factory: endless robot factory generated live in-browser with WFC and OpenUSD — mishig25 · 2026-11-07
- Real-world robot evals go open on Hugging Face for the community — DJiafei · 2026-11-07
- Playing rock-paper-scissors with a robot next to a flying car: 'We live in the future we were promised' — beffjezos · 2026-11-07
- Darwin Robotics Raises $3.6M Pre-Seed to Build Robot Technicians for Data Centers — ycombinator · 2026-11-07
- Unitree built its own automation lines — and cheap Chinese labor means it barely needs them — teortaxesTex · 2026-11-07
- Hugging Face hosts ~100K open robotics datasets, up 48% in six months — micoolcho · 2026-11-07