Inspiration
Low Earth orbit is more congested than ever, with thousands of active satellites and debris objects generating constant conjunction alerts. Current workflows require human analysts to manually investigate high-risk events. We wanted to build an autonomous on-device agent that could reduce that operational burden while keeping sensitive decision-making local using OpenClaw, NVIDIA Nemotron, and DGX Spark.
What it does
Orbital is an autonomous conjunction triage agent for satellite operations. It ingests live TLE data, propagates satellite orbits with SGP4, detects close approaches, evaluates collision probability, checks space weather conditions, and drafts maneuver recommendations for human approval. The system runs fully locally on NVIDIA DGX Spark hardware using NVIDIA Nemotron models orchestrated through OpenClaw.
How we built it
We built Orbital using:
FastAPI backend for APIs and orchestration React + Three.js frontend for the 3D orbital visualization SGP4 propagation and conjunction screening for deterministic orbital calculations SQLite persistence for event and verdict memory NVIDIA DGX Spark for local edge AI compute NVIDIA Nemotron running locally on DGX Spark OpenClaw for autonomous agent orchestration, reasoning loops, and tool calling MCP tools for querying conjunctions, memory, propagation, and maneuver simulation
Challenges we ran into
One of the biggest challenges was balancing deterministic aerospace calculations with autonomous AI reasoning. We learned that Nemotron should never perform orbital mechanics itself and instead reason over trusted tools orchestrated through OpenClaw. We also had to optimize agent latency on DGX Spark, reduce redundant tool calls, and build realistic conjunction scenarios that would survive fresh propagation checks.
Accomplishments that we're proud of
We successfully built an end-to-end autonomous investigation loop running fully locally on DGX Spark. Orbital can detect a conjunction, investigate it through OpenClaw tool orchestration, reason about operational risk with Nemotron, and generate structured maneuver recommendations without relying on cloud inference. We are especially proud of integrating OpenClaw, local Nemotron inference, DGX Spark edge compute, persistent operational memory, and a live orbital visualization into a cohesive system.
What we learned
We learned how autonomous agents can operate beyond chat interfaces by coordinating tools, memory, and structured workflows through OpenClaw. We also gained experience with orbital mechanics, conjunction analysis, local NVIDIA Nemotron inference, and secure edge-based AI systems running directly on DGX Spark hardware.
What's next for Orbital
Next, we want to improve maneuver optimization, add support for higher-fidelity orbital propagators, integrate private ephemeris sources, and expand the multi-agent workflow for large satellite constellations. We also want to strengthen the NemoClaw security layer and support real-time autonomous monitoring at larger orbital scales.
Log in or sign up for Devpost to join the conversation.