Visual MuJoCo simulation where a Hunter SE robot navigates a warehouse. 6 dataflows demonstrate progressive octos integration from pure nav to obstacle avoidance.

You're building a warehouse patrol robot. Your pipeline works in the mock examples (01-04), but you need to answer harder questions: Does the path planner actually avoid obstacles? Does the Pure Pursuit controller track the path within 0.5m? Does the robot arrive at station A before the 120s deadline? You can't answer these with a mock that just returns "Arrived at A" after 0.5 seconds.

Mock robots test your agent logic. Simulation tests your whole stack: physics, control, planning, perception, and agent logic together. The gap matters:

• A mock navigate_to always succeeds. In simulation, the robot can overshoot, oscillate, or stall — exactly like real hardware.
• A mock returns instantly. Simulation reveals timing: does the planner keep up at 100ms? Does the LLM respond before the robot reaches the obstacle?
• A mock has no spatial awareness. Simulation shows: is the robot actually at y=10.0 when it claims to be at station A?

This example runs the same 5 octos patterns (pipeline, replan, safety, cyclic) but with a real physics loop — MuJoCo at 100Hz, Frenet planning at 10Hz, Pure Pursuit control at 100Hz, Rerun visualization at 10Hz.

• Full dora-nav pipeline ported to Python (5 C++ nodes → Python)
• MuJoCo physics simulation with Rerun 3D visualization
• Octos pipeline patrol (deterministic, no LLM)
• Octos LLM replanning (obstacle handling)
• Octos safety tiers (observe-only mode)
• Octos cyclic patrol (continuous loops)

pip install dora-rs pyarrow numpy mujoco rerun-sdk

You also need:

• dora-mujoco: symlink to dora-moveit2/dora-mujoco (MuJoCo sim node)
• Model meshes: symlink models/hunter_se and models/GEN72 to mesh directories
• Ollama (for LLM examples): ollama pull qwen3:30b

cd 05-slam-nav-sim

# Symlink dora-mujoco
ln -sf /path/to/dora-moveit2/dora-mujoco dora-mujoco

# Symlink model meshes
ln -sf /path/to/dora-moveit2/examples/hunter_with_arm/models/hunter_se models/hunter_se
ln -sf /path/to/dora-moveit2/examples/hunter_with_arm/models/GEN72 models/GEN72

export MUJOCO_GL=cgl  # macOS
dora up && dora start dataflow_nav_sim_py.yaml --attach

dora up && dora start dataflow_octos_nav.yaml --attach

dora up && dora start dataflow_octos_replan.yaml --attach

Uses octos serve as the agent brain with the full Rust octos-agent crate (hooks, MCP, memory, compaction). The MCP bridge connects octos to the dora dataflow.

./start_octos_serve.sh
# Opens http://localhost:3141 — send commands via web dashboard
# Or: octos chat (in another terminal)

octos serve (Rust agent + LLM)
    ↕ MCP stdio protocol
mcp_dora_bridge.py (tool forwarder)
    ↕ Unix socket
dora dataflow (nav-bridge + sim + planning)

mujoco-sim → pose-extractor → road-lane-pub ← pub-road
                                    ↓
                               planning ← task-pub-stub
                              ↙        ↘
                        lat-control  lon-control
                              ↘        ↙
                             nav-bridge ←→ robot-edge-a (optional)
                                 ↓
                            wheel_commands → mujoco-sim

                            rerun (3D viz)