SJTU's MAVLN: first systematic multi-agent vision-language navigation benchmark with 11,724 episodes
SJTU · hf · 2026-10-13
Shanghai Jiao Tong University presents MAVLN, the first systematic formalization of multi-agent vision-and-language navigation as a constrained coordination problem: 11,724 episodes across 145 scenes with up to four agents. They also introduce TRISS, coupling an LLM subtask scheduler, shared topological memory, and conflict-aware execution, establishing a baseline that reveals substantial room for improvement in scheduling, planning, and execution.
More from Embodied
- Lightspeed maps robotics landscape on two axes; sees DeployCos as the emerging opportunity layer — FinanceYF5 · 2026-10-13
- The modern robotics infrastructure stack: 16 layers from sim to agents — FinanceYF5 · 2026-10-13
- Lamport Robotics shows off big workshop robot Pincer1 — omnivaughn · 2026-10-13
- NVIDIA open-sources Apache-2.0 CAD kernel usd-brep, shaking commercial CAD vendors — yacineMTB · 2026-10-13
- NewEyes opens preorders for $199 AI glasses, $149 camera earbuds, $99 AI pin — omarsar0 · 2026-10-13
- Dev ships 'EveryDay Carry' firmware for AI Passport, hailed as best network radio — ezshine · 2026-10-13