AVALON
A zero-shot indoor navigation framework that leverages human priors, such as floor plans and place cues, to build a hierarchical topological memory, enabling a robot to navigate new environments without prior mapping or on-site training.
Timeline (last 90 days)
AVALON is a zero-shot indoor navigation framework that requires no prior mapping or site training, targeting scenarios where "humans are familiar and the robot is unfamiliar."
AVALON consists of three core components: human-seeded hierarchical spatial memory, node-verified topological execution, and evidence-gated human-robot memory collaborative updating.
AVALON's deployment conditions require only a floor plan, some location cues, and one initial topology-odometry alignment, with no need for prior on-site traversal or site-specific model training.
The topological node network is the core abstraction of AVALON's spatial representation, retaining only key locations (nodes) and their connectivity relations (edges), with no need for continuous coordinates.
Translation re-anchoring is the key operation by which AVALON maintains a sense of position without a global SLAM map, namely that whenever the robot confirms arrival at a verified topological node, it re-anchors the origin of the odometry coordinate frame to that node's known coordinates.
Evidence gating is a memory-write admission mechanism in which memory is updated only when a human-reported change is relevant to the current task and the robot perceives corresponding physical evidence on site.
AVALON was experimented on a Unitree Go2 quadruped robot, completing a 12.81-meter task from Intersection to Dining together with the X-NavDP policy.
AVALON together with X-NavDP took 46.4 seconds on this task, succeeding in all five runs (5/5), with no mapping required throughout.
AVALON's robot execution time was 40.9% less than DimOS-Go2's.
AVALON replaces a posteriori, geometric dense maps (robot-built SLAM) with a priori, semantic sparse structure (human-seeded topological memory); the cost is losing the precise local planning capability brought by continuous metric maps, and the benefit is zero-shot deployment and faster task startup.
2 more timeline events
All Facts (12)
AVALON is a zero-shot indoor navigation framework that requires no prior mapping or site training, targeting scenarios where "humans are familiar and the robot is unfamiliar."
50%UnverifiedAVALON consists of three core components: human-seeded hierarchical spatial memory, node-verified topological execution, and evidence-gated human-robot memory collaborative updating.
50%UnverifiedAVALON's deployment conditions require only a floor plan, some location cues, and one initial topology-odometry alignment, with no need for prior on-site traversal or site-specific model training.
50%UnverifiedThe topological node network is the core abstraction of AVALON's spatial representation, retaining only key locations (nodes) and their connectivity relations (edges), with no need for continuous coordinates.
50%UnverifiedTranslation re-anchoring is the key operation by which AVALON maintains a sense of position without a global SLAM map, namely that whenever the robot confirms arrival at a verified topological node, it re-anchors the origin of the odometry coordinate frame to that node's known coordinates.
50%UnverifiedEvidence gating is a memory-write admission mechanism in which memory is updated only when a human-reported change is relevant to the current task and the robot perceives corresponding physical evidence on site.
50%UnverifiedAVALON was experimented on a Unitree Go2 quadruped robot, completing a 12.81-meter task from Intersection to Dining together with the X-NavDP policy.
50%UnverifiedAVALON together with X-NavDP took 46.4 seconds on this task, succeeding in all five runs (5/5), with no mapping required throughout.
50%UnverifiedAVALON's robot execution time was 40.9% less than DimOS-Go2's.
50%UnverifiedAVALON replaces a posteriori, geometric dense maps (robot-built SLAM) with a priori, semantic sparse structure (human-seeded topological memory); the cost is losing the precise local planning capability brought by continuous metric maps, and the benefit is zero-shot deployment and faster task startup.
50%UnverifiedNode verification requires that when the robot arrives somewhere, it must confirm that it has indeed reached the expected node; otherwise topological reasoning will cascade into errors.
50%UnverifiedThe experimental scale that AVALON has currently made public is limited, and validation on a single task and a single robot platform is not yet sufficient to demonstrate the framework's generalization ability in more complex scenarios.
50%