Internship Progress Week 6

Internship Progress Week 6

This week, I updated the expert navigation setup, collected a new dataset and trained two versions of the local trajectory prediction network.

Navigation setup

I reorganized the project into ImitationLearning4LocalPlanner/ros2_ws, updated the launch files and documented the ROS 2 Lyrical, Gazebo and Pixi setup.

The navigation configuration now uses longitudinal motion and rotation, with no lateral velocity commands (vy = 0). After testing DWB and Rotation Shim, I switched to Regulated Pure Pursuit (RPP) to address the observed tracking and orientation problems. NavFn with Dijkstra remains the global planner.

The main settings are:

• Desired forward velocity: 1.0 m/s, with reversing disabled in RPP.
• Lookahead distance: 2.0 m, with rotation-to-heading enabled.
• Goal tolerances: 0.3 m and 0.2 rad, with path_length_tolerance set to 2.0 m.
• PoseProgressChecker, which accounts for rotation as well as translation.

I also corrected the odometry-to-TF bridge, removing an extra 90-degree rotation and consistently using base_link in Nav2. This improved the alignment between LiDAR observations and the map. Recovery commands were remapped to /autonomous_car/cmd_vel using Twist, and a backup manoeuvre was successfully tested. Reversing remains available for recovery behaviours.

New dataset

With the corrected setup, I recorded 10 navigation episodes. The processing script removes waiting periods between goals, excludes problematic sections in episodes 3 and 5, and rejects samples with stale or incomplete inputs. A sample is also excluded if its future trajectory crosses a cut or an episode boundary.

The resulting dataset contains 4,480 samples, sampled at 5 Hz:

Training	Validation	Test
3,055	735	690

The split is made by episode, and normalization statistics are calculated only from training data.

Each input contains LiDAR ranges, ten relative global-plan waypoints and the measured velocities (v, w), where v = vx. The output contains five relative (x, y) points at 1, 2, 3, 4 and 5 seconds, reconstructed from future odometry. The network predicts a local trajectory rather than actuator commands.

Network training and results

I implemented an adaptation of the local planner from Pokle et al., Deep Local Trajectory Replanning and Control for Robot Navigation. It combines separate LiDAR, global-plan and odometry branches through attention. The pedestrian branch and learned velocity controller are omitted, and the output contains only (x, y) coordinates.

In the first version, attention collapsed almost entirely onto LiDAR. For v2, I added LayerNorm to the three feature vectors and initialized attention uniformly. I trained both versions on my GPU, selecting the checkpoint with the lowest validation MSE.

Model	Validation ADE	Validation FDE	Episode 10 ADE	Episode 10 FDE
v1	0.259 m	0.493 m	0.388 m	0.659 m
v2	0.120 m	0.202 m	0.240 m	0.406 m

ADE measures the average distance error across the five predicted points; FDE measures the error at five seconds. The best v2 checkpoint was saved at epoch 20. Its validation attention weights averaged approximately 65% LiDAR, 33% global plan and 3% odometry, so attention no longer concentrated exclusively on LiDAR. These weights alone do not establish how much each input affects predictions.

Next steps

Next, I will check whether changing the global plan and velocity inputs meaningfully changes the predictions, collect new held-out routes, and test the predicted trajectories with a conventional controller in simulation.

The current results are offline measurements from one map. Since episode 10 has already been inspected for both versions, a fresh test set is needed for an independent final evaluation. Closed-loop tests will assess goal completion, collisions and blocking behaviour.