F1Tenth Autonomous Racing

Overview

Team 10's F1TENTH car is a 1/10-scale autonomous race car built on a Traxxas Slash 4×4 with a Jetson Orin Nano, a Hokuyo 2D LiDAR and a VESC motor controller, running ROS 2 Humble. It started as a course build in CMU 16-663, raced at two Roboracer competitions in 2026, and the work continues in Prof. John Dolan's AIMS Lab.

The car

The stock Traxxas chassis carries a two-level deck. The Hokuyo LiDAR sits at the front on the center line, the Jetson Orin Nano in the middle, and the power distribution board, VESC and Wi-Fi antennas toward the rear, with the LiPo lead on an XT60 connector.

Timeline

1. Course build, CMU 16-663 (Spring 2026). Built the full autonomy stack on the car as Team 10.
2. 28th Roboracer Autonomous Racing Competition, IEEE IV 2026, Detroit. Finished 5th of 17 teams with a Frenet-frame Pure Pursuit controller and a LiDAR avoidance layer.
3. 31st Roboracer Autonomous Racing Competition, IROS 2026, Pittsburgh. Rebuilt the race stack around map-generated racelines, Barrier State Control for obstacle avoidance and fused localization, building on a barrier-state approach proposed by a collaborator. An Economic MPC with barrier states is in development as the successor controller.
4. AIMS Lab research (ongoing). Continuing the platform under Prof. John Dolan.

Autonomy stack

• Reactive avoidance: Follow the Gap for obstacle avoidance.
• Localization and mapping: scan matching, particle filter localization and graph-based SLAM.
• Planning: RRT* path planning in unknown environments.
• Control: Pure Pursuit path tracking with tuned PID for high-speed trajectories.
• State estimation: IMU, wheel odometry and 2D LiDAR fused for robust state estimates.
• Perception: monocular depth estimation for navigation when sensing is degraded.

Approach

Three controller tiers

• Baseline: Frenet Pure Pursuit with a model-based speed profile and barrier-state avoidance.
• Near Star: a Frenet Economic MPC that rewards progress and lets a friction-ellipse constraint set corner speed, so late apexes and trail braking come from the cost rather than hand-written rules.
• Blue Sky: adaptive MPC with learned weights or a distilled policy.

Racelines from the map alone

Hand-logged waypoints added curvature noise, so a generator builds three lines from the SLAM map: stable, minimum curvature, and minimum lap time with CasADi and IPOPT. On the IROS organizer map it extracted a 208.7 m loop with widths of 1.68 to 3.58 m.

Car data over simulator defaults

Settings tuned in the gym did not transfer: its tire model gave an understeer gradient about 7× smaller than the value tuned on the car. Bag analysis found about 140 ms between the drive command and the real yaw rate, which explained a 1.3 Hz weave. The car became the ground truth, with one change per run.

Fused localization

In one wide corner the particle filter lost heading while staying confident. A localization package now gates particle filter poses against a map-frame EKF and switches between fused, dead-reckoning and stop modes with speed limits.

Results

• 5th of 17 teams at IEEE IV 2026, Detroit.
• New Baseline on the car: 2 clean laps at 57.3 s, 0.18 m RMS lateral error, steering at its limit 1.7% of the time.
• Avoidance in a car-calibrated simulation: 111 of 112 stress runs clean after fixing the side latch; phantom detections on real scans cut from 25 to 8.
• Localization replay on real bags: heading error during the corner loss fell from 78° to 14°, and from 180° to 12° in the run that spun.

Lessons learned

• A derivative gain can hide feedforward. KD was quietly encoding curvature feedforward, which caused a see-saw between tuning straights and tuning curves until the feedforward term was made explicit.
• Fixed gains fail at speed. Speed-blind gains went unstable at high speed, and steering saturation plus low-speed gain amplification caused the inner-wall crashes in U-turns.
• Match the track, not the default. The avoidance half-width had to match the real track (about 0.50 m on a 1 m track, not 0.80 m).
• Some VESC problems are not tunable. "Allow Braking" in the speed PID produced a confirmed limit cycle, and v6 configuration files silently dropped parameters on v7.00 firmware.
• ROS 2 fails quietly. A node name that does not match its YAML key binds no parameters and raises no error.

What's next

Extending the stack toward an imitation-learning architecture that distills expert driving behavior into the planner, for faster onboard reaction at higher speeds. Nearer term:

• Near Star on acados, with the Baseline as a hot standby behind a drive arbiter.
• System identification on the car, then a dynamic single-track model with Pacejka tires.
• Delay compensation and moving the controllers onto the fused pose.