Gb223/autonomous-rl-mpc-usv

🛥️ USV Safe RL Workspace

Shielded Reinforcement Learning for Unmanned Surface Vehicles

A hybrid PPO + MPC safety-filter architecture for autonomous marine navigation, trained in a custom Gymnasium environment and validated in Gazebo 3D (ROS 1).

Research internship — State Key Laboratory of Ocean Sensing, Zhejiang University · Polytech Nice Sophia (Université Côte d'Azur)

---

Overview

Deep Reinforcement Learning produces agile but unverifiable navigation policies. Model Predictive Control produces safe but conservative (and expensive) ones. This project synthesizes both:

The RL agent (PPO) acts as the brain — high-level, adaptive navigation. The MPC Predictive Safety Filter acts as the spinal cord — a real-time reflex enforcing hard kinematic constraints on every command before it reaches the thrusters.

The full pipeline goes from a pure mathematical prototype (10 000+ simulated steps/second) to a physical 3D deployment inside Gazebo with a custom multi-layer LiDAR, wind disturbances and wave physics.

                ┌──────────────────────────────────────────────────────────┐
   LiDAR 3×360° │                                                          │
   + Odometry ──┼──▶ Observation (18-D) ──▶ PPO Policy ──▶ u_RL            │
                │                                            │             │
                │                                            ▼             │
                │                         ┌─────────────────────────────┐  │
                │                         │  MPC SHIELD (QP, SLSQP)     │  │
                │                         │  Fossen 3-DOF prediction    │  │
                │                         │  d_safe = 2.0 m, H = 2 steps│  │
                │                         └──────────────┬──────────────┘  │
                │                                        ▼                 │
                │            Emergency reflex (<2.5 m) ─▶ u_safe ──▶ Thrusters (PWM)
                └──────────────────────────────────────────────────────────┘
                                  single synchronous 10 Hz control loop

---

Demonstrations

3D validation runs in Gazebo. Drop the corresponding .gif files into report_files/ using the exact names below and they will render automatically.

Test 1 — Sanity check (chicane)	Test 2 — Narrow corridor
Smooth Sim-to-Sim transfer, shield rarely triggered	±30° heading clamp keeps the trajectory linear in an 8 m channel

Test 3 — Close quarters (orthogonal avoidance)	Test 4 — The Storm (lateral wind)
3-layer LiDAR fusion resolves overlapping buoys during a 90° turn	The agent adopts a "crab" sideslip; the shield governs lateral drift

Test 5 — The Full Gauntlet (stress test)
Longest course, tightest geometry, heaviest logging load — zero collisions

---

Method

Vehicle dynamics — Fossen 3-DOF

The MPC predicts the vessel using the standard maneuvering model with pose $\eta = [x, y, \psi]^T$ and body velocities $\nu = [u, v, r]^T$:

$$\dot{\eta }=R(\psi )\nu ,M\dot{\nu }+C(\nu )\nu +D(\nu )\nu =\tau \backslash dot\{\backslash eta\}\; =\; R(\backslash psi)\backslash ,\backslash nu,\; \backslash qquad\; M\backslash dot\{\backslash nu\}\; +\; C(\backslash nu)\backslash nu\; +\; D(\backslash nu)\backslash nu\; =\; \backslash tau$$

where $M$ includes added mass, $C(\nu)$ the Coriolis terms, $D(\nu)$ the non-linear hydrodynamic damping, and $\tau = [\tau_u, 0, \tau_r]^T$ the differential thrust of the two motors.

The MPC safety shield

At each tick, the RL action $u_{RL}$ is projected onto the safe set by a Quadratic Program (scipy.optimize.SLSQP) over a 2-step horizon ($dt = 0.5$ s ⇒ 1 s look-ahead):

$$\underset{u_{MPC},ϵ}{\min }\frac{1}{2}\mathrm{\parallel }{u}_{MPC}-u_{RL}{\mathrm{\parallel }}_R^2+W_{ϵ}{ϵ}^2\backslash min\_\{u\_\{MPC\},\backslash ,\backslash epsilon\}\; \backslash tfrac\{1\}\{2\}\backslash |u\_\{MPC\}\; -\; u\_\{RL\}\backslash |\_R^2\; +\; W\_\backslash epsilon\; \backslash epsilon^2$$

subject to the Fossen dynamics, actuator saturation, and the collision constraint $|P_{USV}(t+H) - P_{obs}(t+H)| \ge d_{safe} - \epsilon$. The heavily penalized slack $\epsilon$ guarantees feasibility under severe disturbance: instead of an infeasible solve, the filter degrades gracefully into maximum-effort emergency braking.

Curriculum learning (10 scenarios)

Training PPO directly on dynamic traffic collapses the policy, so the agent is grown through a weighted curriculum:

#	Scenario	Skill taught
1	free	Locomotion, heading correction, stopping precision
2	static_single	Elementary avoidance without over-deviating
3	static_multi	Chicane anticipation, multi-obstacle planning
4	narrow_corridor	Precision navigation in an 8 m channel
5	moving_follow	Safe overtaking with lateral offset
6	moving_cross	COLREG-like crossing: cross or yield
7	moving_target	Pursuit of a mobile waypoint
8	fast_unpredictable	Robustness / spatial margin against erratic traffic
9	variable_distance_world	Geometric generalization (30 m → 100 m)
10	mixed_dynamic	Final synthesis exam

Key stabilizers: scenario sub-weighting against catastrophic forgetting, learning rate $3\times10^{-5}$, PPO clip range $0.1$, spatial jitter / domain randomization, and a shield-intervention penalty $-0.1,|u_{MPC}-u_{RL}|_2$ that forces the network to internalize the physical constraints instead of leaning on the filter.

---

Results

Final converged policy, 100 episodes per scenario:

Scenario	Success	Collision	Timeout	Mean length
free	1.00	0.00	0.00	189.5
narrow_corridor	1.00	0.00	0.00	173.3
moving_follow	1.00	0.00	0.00	161.4
variable_distance_world	1.00	0.00	0.00	174.2
static_multi	0.98	0.02	0.00	140.0
moving_cross	0.96	0.04	0.00	168.1
moving_target	0.96	0.00	0.04	373.5
static_single	0.94	0.01	0.05	203.4
mixed_dynamic	0.91	0.09	0.00	142.1
fast_unpredictable	0.88	0.12	0.00	169.4

Network health (TensorBoard): approx_kl ≈ 0.006 · explained_variance = 0.925 · value_loss = 0.032 · clip_fraction = 0.05 · std = 1.13 — a well-behaved trust region, an accurate critic, and preserved exploration.

Two engineering findings worth highlighting

• Over-engineering is the enemy. The first Gazebo deployment chattered and pirouetted. Adding defensive layers (state caches, async callbacks, nested clamps) made it undebuggable. The fix was removal: a single synchronous 10 Hz loop where observation, inference, shield and actuation share one atomic clock.
• The Observer Effect. Running rosbag record collapses the Gazebo Real-Time Factor, desynchronizing the closed loop and making trajectories jagged. Unrecorded runs are perfectly smooth. The degradation became a free stress test: even with a severely lagging RL loop, the shield kept the collision count at zero.

---

Getting started

1. Requirements

• Ubuntu 20.04 + ROS Noetic (ros-noetic-desktop-full) and Gazebo 11
• Python 3.8+ with numpy, scipy, torch, stable-baselines3, gymnasium, matplotlib

A ready-to-use dev container is provided (osrf/ros:noetic-desktop-full + noVNC desktop on port 6080):

pip3 install numpy scipy torch stable-baselines3 gymnasium matplotlib

2. Build the ROS workspace

catkin_make          # or: catkin build
source devel/setup.bash

3. Run the 3D simulation

./e1_myboat_gazebo_view1.bash      # ocean world 1 (worlds/myocean1.world)
./e1_myboat_gazebo_view2.bash      # ocean world 2
./e1_myboat_gazebo_view3.bash      # ocean world 3
./e2_myboat_manual_control.bash    # manual teleoperation of the thrusters
./test_py_myboat_control.bash      # Python control node smoke test

Then start the shielded controller (loads usv_rl_brain_final.zip + vecnormalize_final.pkl):

rosrun myboat_gazebo usv_controller_final.py     # or: python3 clean_version/usv_controller_final.py

4. Train / evaluate the policy (no Gazebo required)

cd clean_version
python3 train_rl.py                 # PPO curriculum training
python3 resume_train.py             # fine-tune an existing checkpoint
python3 casino_method.py            # random-restart search on the 10-scenario mix
python3 evaluate_each_scenario.py   # per-scenario success / collision / timeout rates
python3 plot_metrics.py             # learning curves

5. Reproduce the 3D test figures

cd tests_file
python3 plot_rosbag.py              # trajectory + thruster command plots from a .bag file

---

Documentation

The complete internship report (theory, training methodology, Sim-to-Sim transfer, validation campaign, LiDAR URDF appendix) is available in report_files/, together with the intermediate phase reports.

---

Future work

The architecture is hardware-ready. The next milestone is deployment of the containerized controller onto the physical USV for real lake trials — where the MPC shield and the domain randomization used during training are expected to absorb the remaining Sim-to-Real gap without risky on-site retraining.

---

References

1. Wabersich, K. P., & Zeilinger, M. N. (2021). A Predictive Safety Filter for Learning-Based Control of Constrained Nonlinear Dynamical Systems. Automatica, 129, 109597.
2. Romero, A., Song, Y., & Scaramuzza, D. Actor-Critic Model Predictive Control: Differentiable Optimization Meets Reinforcement Learning for Agile Flight. IEEE T-RO.
3. Song, Y., & Scaramuzza, D. (2022). Learning High-Level Policies for Model Predictive Control. IEEE RA-L.
4. Brunke, L., et al. (2022). Safe Learning in Robotics: From Learning-Based Control to Safe Reinforcement Learning. Annual Review of Control, Robotics, and Autonomous Systems, 5, 411–444.
5. Fossen, T. I. (2011). Handbook of Marine Craft Hydrodynamics and Motion Control. John Wiley & Sons.

---

Acknowledgments

Work carried out at the State Key Laboratory of Ocean Sensing, Zhejiang University, under the supervision of Prof. Weichang Li and Dr. Huarong Zheng, with the academic support of Prof. Noëlle Stolfi (Polytech Nice Sophia). The baseline simulation assets are derived from the open-source VRX project.

_{Author: Gabriel Ferré — 4th year, Applied Mathematics & Modelling}