You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
This repository was archived by the owner on Apr 10, 2026. It is now read-only.
ERIC is a tracked ground robot that runs NVIDIA Cosmos Reason 2 fully at the edge — no cloud, no server, no internet. A plain-English mission briefing is all it needs. Cosmos handles everything else: parsing multi-step goals, navigating via live camera frames, scanning 360° across 42 positions, escaping obstacles, conducting character conversations, and confirming targets before triggering alerts.
Cosmos is not just the object detector — it is the entire decision layer. Navigation commands, obstacle escape angles, eye-contact gating, false-positive rejection, and mission completion announcements all flow through a single 2B VLM running at 16–17 tokens/sec on a $250 Jetson Orin Nano.
The demo tests open-world object search: ERIC locates a target specified only by natural language description — no pre-trained classifier, no environment map. A size- and colour-matched decoy is placed in plain sight with no briefing hint; Cosmos must discriminate on shape and feature reasoning alone. The target is partially occluded and wrapped in reflective material.
Hardware total: under $800 CAD. Built solo in 3 days.