For full documentation please visit the docs site: https://airou-lab.github.io/general_wiki_website/

To clone run:

git clone --recurse-submodules -j8  git@github.com:airou-lab/arcpro_system.git

(To add build passing tags)

(To add current version deployment)

(To add current maintainer)

Maintainer until Dec 2027: arikak@ou.edu

When building on new device run in project root:

rosdep install --from-paths src -y --ignore-src

Running Examples:

# waypointer example:
 ./waypoint.sh 
 # RL example:
./arcpro_rl.sh 

Reinforcement Learning (Isaac Lab - Phase 3)

The ARCPro RL system is currently migrating to Isaac Lab to enable massive parallel training and resolve simulation stability bottlenecks.

• Status: Phase 3 Migration in Progress.
• Goal: Vectorized training with 128+ agents using the Manager-Based RL API.
• Digital Twin: Verified metric asset (25cm WB, 4.092kg) calibrated for high-fidelity physics.

Simulation & Hardware Alignment (Digital Twin)

The simulation model has been meticulously calibrated to match ARCPro hardware for Zero-Shot Sim2Real transfer.

The F1Tenth simulation model in src/examples/ARCPro_RL/arc_rl_isacc_sim/f1tenth_trainer/assets/F1Tenth.usd has been refactored to match exact ARCPro hardware specifications for Zero-Shot Sim2Real transfer.

Hardware Specifications Applied

• Total Weight: 4092 g (4.092 kg)
• Wheelbase: 25.0 cm (Kinematic length between axles)
• Track Width: 24.0 cm (Horizontal width between wheels)
• Wheel Radius: 5.0 cm
• LIDAR Offset: [235, 0, 265.23] mm (Forward, Left, Up from chassis root)
• Camera Offset: [145, 0, 195] mm (Realsense D435 centered)

Stability Fixes (Isaac Sim 2025)

To resolve physics explosions and high-frequency jitter (NaN errors) common with 34-joint high-fidelity models in Isaac Sim 2025, the following "Runtime Stability Pass" is implemented in isaac_direct_env.py:

1. Mass Force-Injection: Explicitly sets rigid body masses (3.0kg chassis, 0.15kg wheels) at environment startup to eliminate zero-mass errors.
2. Damping Override: Bypasses deprecated USD attributes by forcing 50.0 damping and 1000.0 stiffness across all 34 joints via the ArticulationView API.
3. Solver Beefing: Increases physics precision to 64 position iterations and 32 velocity iterations to eliminate mathematical drift and stabilize the 26-joint suspension.

Vision & Policy Stability

The following refinements were applied to ensure reliable data flow and training convergence:

1. Vision Verification: Confirmed stable frame capture via IsaacDirectEnv. Relocated default spawn to (-125.0, 62.0) to ensure the robot always starts on high-contrast textured track segments, resolving "black-screen" initialization bugs.
2. Policy Gradient Patch: Patched HierarchicalPathPlanningPolicy to recreate the optimizer after custom hierarchical heads (Planning/Control) are initialized. This ensures all sub-networks are registered for backpropagation, which was previously blocked by the standard SB3 super().__init__ sequence.
3. Single-Process Direct API: Refactored train_direct.py to initialize SimulationApp before any other imports. This prevents GPU context collisions and ensures the Direct API maintains a stable singleton connection to the physics engine.

to just view the xacro file, run

./src/base/robot/urdf/models/rsp_xacro_test.sh

# AND in host terminal for gui
ros2 run joint_state_publisher_gui joint_state_publisher_gui