Skip to content

sim2real.md

Latest commit

 

History

History
44 lines (36 loc) · 2.33 KB

sim2real.md

File metadata and controls

44 lines (36 loc) · 2.33 KB
Raw

Sim-to-real and ROS 2

The latest ROS pair is supplied in models/: sim2real_nominal_policy.pt and sim2real_adversarial_policy.pt, each with its matching *_norm_stats.npz. These are the user-confirmed robot models. No robot recordings are bundled.

Training randomizes 85% of rollouts: radius/base scales 0.94–1.06, independent wheel gains 0.88–1.12, motor time constant 0.02–0.12 s, action delay 0–3 steps, observation delay 0–2 steps. State noise and DART recovery collection are also enabled. These are synthetic training ranges, not measured robot calibration.

evaluate_sim2real_interpolated_wall.py evaluates newly trained model files in artifacts/ by default. To evaluate the supplied pair, first copy the four models/ files to a newly created artifacts/ directory. Do not overwrite a training run you want to retain. --include-legacy requires the older checkpoints.

Robot node

In a sourced ROS 2 workspace with colcon, place ros2_nn_trajectory_tracker/ inside the workspace's src/ directory. Its Python must have NumPy and PyTorch. Build and source the workspace:

colcon build --packages-select nn_trajectory_tracker
source install/setup.bash
ros2 launch nn_trajectory_tracker lemniscate_tracker.launch.py \
  model_path:=/absolute/path/to/paper_companion/models/sim2real_nominal_policy.pt \
  norm_stats_path:=/absolute/path/to/paper_companion/models/sim2real_nominal_norm_stats.npz \
  max_linear_velocity:=0.2 max_angular_velocity:=0.8

Select the adversarial model by changing both paths. Paths are explicit launch arguments; model files need not be copied into the ROS package. Defaults are /odemetry/filtered (spelling retained from the robot interface), /cmd_vel, vicon/world, CPU and 60 Hz. Override topic/frame launch arguments to match the robot. Check odometry, coordinate conventions, wheel ordering, command scaling and stale-odometry behavior before enabling motion.

The node converts predicted right/left wheel speeds to linear/angular velocity. It receives eight raw features and appends heading trig features internally. No inference-time smoothing is added. The package supplies clean policy tracking; it does not implement the real-time UAP/FGSM experiment harness. Attacks in this release run in simulation. ROS build, middleware and physical motion require validation on the target machine.