AGENTIC ROBOT NAVIGATION

Set the goal. Let your robot search.

Describe what you want to find. Nova-HEI observes, decides where to explore, and guides the robot with vision-language navigation.

Modular architecture, clean boundaries

One agent. Independent adapters for navigation models and execution backends.

Engineering interface overview: one unchanged Agent connects to independent navigation model and execution backend adapters. LightNav, Go2 and the MuJoCo simulation are built-in implementations. Dashed Custom slots indicate extension points, not connected or validated implementations. The LightNav neural-network symbol represents a learned navigation model, not its actual topology. Lines are structural relationships, not execution order.

Runtime workflow overview

Harness supports continuous navigation, timely adjustment and completion checks through in-task memory, monitoring and scheduling.

From route execution to autonomous search

Memory

Observations, routes and action history, recalled within a task.

Monitor

A monitor watches for target cues during navigation and triggers a pause for the agent to review.

Scheduling

VLN keeps moving; the agent replans when progress stalls.

From a goal to exploration

Often, we know what we want to find, but not where it is. Nova-HEI lets robots take on that search. Describe your goal in everyday language, and the robot explores its surroundings without needing you to specify a location or guide it at every turn.

The target may be in another room or around a corner, out of sight. The robot keeps your request in mind as it searches, follows clues it discovers along the way, and approaches possible matches to check whether they meet your requirements. You give it a complete search task, without having to issue a new instruction for each move.

See it in action

MuJoCo

Method comparison

Go2

Real-world deployment

Benchmarks

Habitat-GS

SR (%) ↑
SPL ↑
NE (m) ↓
OS ↑
nDTW ↑

HM3Dv2

SR (%) ↑
SPL ↑
NE (m) ↓
OS ↑

HM3D-OVON

SR (%) ↑
SPL ↑
NE (m) ↓
OS ↑