Point Nemo: ARIA Autonomous Mars Rover

https://docs.google.com/document/d/1cEeME2bTkU9q_nxCBheeANVDT9DVEgIYikh7KD2irq8/edit?tab=t.0 Project for the Mars Rover Hackathon using NemoClaw and OpenClaw on NVIDIA DGX Spark.

Overview

ARIA (Autonomous Rover Intelligence Architecture) is designed to operate autonomously in the Jezero Crater, overcoming the 3-22 minute communication latency between Earth and Mars. It uses a hierarchical agent structure to orchestrate complex missions like geological discovery and environmental monitoring.

System Architecture

• Supervisor (Main): Orchestrates sub-agents and manages high-level mission objectives.
• Navigation Agent: Handles path planning, hazard avoidance, and locomotion.
• Power Agent: Monitors battery SoC, solar irradiance, and thermal health.
• Science Agent: Prioritizes geological targets and processes vision data.

Workspace Structure

• workspace/identities/: Markdown files defining each agent's persona.
• workspace/souls/: Markdown files defining the autonomous reasoning loops.
• workspace/tools/: Python implementations for rover-facing tools and skill wrappers.
• panda_tests/: Local Panda3D simulator assets plus the source-of-truth seed data used by rover skills in development.
• workspace/policies/: Security presets for sandbox resource access.
• workspace/openclaw_config.json: Configuration patch to register agents and skills.

Deployment Instructions

Prerequisites

• NemoClaw CLI installed on the host.
• A running NemoClaw sandbox (e.g., test-agent).
• jq installed on the host (for configuration patching).

Deploying the Code

Run the sync script from the project root. This script will automatically discover your running sandbox, install the core skills, mirror the panda_tests seed data into /sandbox, patch the OpenClaw configuration, and restart the agent.

See docs/DGX_SPARK.md for the full Spark checklist (vision env, NIM, main-agent science flow).

chmod +x workspace/sync_to_sandbox.sh
./workspace/sync_to_sandbox.sh

Accessing the Rover

Once deployed, shell into the sandbox and start the TUI to interact with ARIA:

nemoclaw <sandbox-name> connect
# Inside the sandbox:
openclaw tui

Vision on DGX Spark (Nemotron Nano VL + science RAG)

• Main agent (Nemotron 30B) reasons and calls tools via rover-tools.
• Vision VL runs inside skills when the agent calls analyze_image_for_science (also classify_terrain_image, capture_image).
• After ./workspace/sync_to_sandbox.sh, edit /sandbox/.openclaw/workspace/vision.env if your VL model id differs from nemotron-nano-vl.
• No SQL database — RAG uses panda_tests/mineral_db/samples.json.
• Local smoke test: python3 -m skills.verify (mock VL). Demo: ./scripts/demo_science.sh.

Science tools available to the main agent

capture_image, classify_terrain_image, analyze_image_for_science, query_mineral_db, log_science_finding, prioritize_targets, schedule_instrument, propose_waypoint, plus navigation/power/comms tools (see docs/tools.md).

Changing Models

You can change the inference model by updating the primary field in workspace/openclaw_config.json and re-running the sync script, or by using the NemoClaw CLI:

nemoclaw inference set --provider compatible-endpoint --model <model-id>

Mission Objectives

1. Survival: Maintain >20% battery and <15° tilt.
2. Discovery: Locate and document a ravine in Jezero Crater.
3. Efficiency: Maximize scientific return within the daily power budget.