ZJU's EvolvingNav predicts where unseen objects moved for persistent robot navigation
ZJU4EmbodiedAI · hf · 2026-10-01
ZJU's 4D Embodied AI team proposes EvolvingNav, addressing Evolving-World Navigation: targets may move while unobserved, making remembered locations unreliable.
- Builds a time-indexed belief from timestamped 3D object histories via a structured persistence-relocation model, distinguishing "still at last observed spot" from "relocated" and keeping probability mass outside known candidates
- An event-driven filter propagates beliefs over time, forecasts target occupancy at inspection times, and fuses new RGB-D evidence; negative observations downweight hypotheses via calibrated visibility-conditioned detection probabilities
- A frozen zero-shot VLM controller acts on the updated belief to replan
- Introduces EvoWorld-Bench: 54 scenes and 803,680 tasks grounded in human activity traces, with controlled changes before and during navigation
Improves navigation success and search efficiency over baselines in both simulation and real-robot experiments.
More from Embodied
- Ex-Amazon warehouse automation lead: Anthropic's robot cost linear projections are naive — BradPorter_ · 2026-10-01
- Carmack: VR pinball on Quest 3 beats real tables, as Pinball FX's Unreal switch costs frames — ID_AA_Carmack · 2026-10-01
- Boston Dynamics gives Atlas 13-DOF dexterous hands built for sim-to-real RL and real work — chris_j_paxton · 2026-10-01
- IROS teardown is still all human labor: when will robots take over setup? — micoolcho · 2026-10-01
- Robot demo wows with tight reactive control, pouring sugar and stirring a glass — chris_j_paxton · 2026-10-01
- DIY robot U-BOT torque test: great on small hills, but overheats on the big one — _Stocko_ · 2026-10-01