AnyGoal: VLM-powered multi-robot navigation sets new SOTA with 52.4% success rate

Researchers from Skoltech have unveiled AnyGoal, a novel multi-agent navigation system that eliminates the need for training and outperforms existing modular pipelines by a wide margin. Traditional end-to-end policies fail in out-of-distribution scenes, while modular systems like GOAT are bottlenecked by closed-set object detectors and heavy 3D memory snapshots. AnyGoal instead places a Vision-Language Model (VLM) at the core of its decision-making. The VLM scores frontier points via a softmax function, and those scores are fused into a shared 2D Gaussian Bayesian Value Map (BVM) that maintains per-pixel posterior distributions (mean and variance) over goal relevance. The BVM is never reset between subtasks, accumulating evidence indefinitely. Frontiers are ranked using a convex blend of the VLM softmax and a Bayesian UCB term from the BVM, while agents are allocated via a greedy dispatcher with spatial-separation penalty and commitment hysteresis—all without a centralized controller.

On the full GOAT-Bench validation unseen split (360 episodes, 2,669 subtasks), a dual-agent AnyGoal system achieves 52.4% Subtask Success Rate at 12.7% SPL—state of the art under the strict physical regime (discrete 0.25m steps, no teleportation, 42° HFOV). This represents a 27.5 percentage point improvement over Modular GOAT (24.9%). Even a single-agent variant reaches 41.9% Subtask SR, demonstrating that the gains stem from the decision architecture rather than the multi-agent setup alone. A four-way perception ablation reveals that replacing closed-set detectors with open-vocabulary ones shifts the dominant failure mode from exploration to goal verification, hinting at future improvement directions. AnyGoal offers a practical path toward lifelong robotic navigation without retraining or extensive memory.