NiFT Autonomous Shuttle thumbnail
NiFT Autonomous Shuttle

Undergraduate Research Winter/Fall 2026 Entrepreneurship Grade: #

kane@umich:~/projects/nift
– ▢ ×
info --detailed
Info           Detail
-------         -------
TIME            Jan 2026 – Present
LAB/ORG         Perot Jain TechLab, Mcity @ U-M; NiFT
ROLE            System Integration Engineer
STACK           Python, C++, ROS 2, CAN bus, Gazebo, Git, Linux
ros2 launch nift_bringup system.launch.py_

Mission

This project aims to develop an autonomous depot shuttle platform for Mcity in collaboration with NiFT. Our goal is to deliver a proof-of-concept transforming the NiFT Shuttle QB into an L4 autonomous vehicle by leveraging infrastructure-based sensing and routing.

Interactive URDF: Click and drag to inspect the shuttle's sensor payload and chassis.

As part of a six-person student team, I architected the end-to-end ROS 2 Humble software stack, bridging high-level autonomy with low-level hardware execution. My core achievements include developing a custom Active Disturbance Rejection Control (ADRC) system alongside an adaptive pure pursuit planner, engineering the SocketCAN communication bridge for by-wire control, and integrating RTK GNSS localization with a fail-safe hardware state machine. To validate our navigation algorithms prior to physical deployment, I also built a high-fidelity Gazebo simulation environment of the Mcity testing grounds, complete with custom URDF models and RViz 2 diagnostic tools.

Implementation

I architected the ROS 2 workspace to achieve the Sense-Think-Act pipeline for the shuttle. For testing, simulation is also developed. The system can be split into 4 core layers:

  • LiDAR-Based Obstacle Detection: Processed raw point cloud data to extract obstacle coordinates and implemented spatial transformations (TF2) to map object detections from the sensor frame to the vehicle's local coordinate system.
  • GNSS-Based State Estimation: Developed an integration for Mcity Octane via Socket.IO to stream global positioning data. Implemented a dual-beacon configuration to accurately compute the shuttle's absolute heading and real-time coordinates.
  • Decision Logic & Safety Layer: Implemented a sensor fusion strategy to synthesize LiDAR and GNSS data, enabling real-time decision-making for path progression and autonomous Emergency Stop (E-Stop) protocols.
  • Global Routing & Waypoint Generation: Engineered a central orchestration node utilizing Lanelet2 HD maps to compute navigable trajectories. Implemented a 3-phase routing strategy (Bezier curve entry, graph search, and exit) to smoothly transition the vehicle between unconstrained off-road poses and structured road networks while strictly adhering to Ackermann kinematic constraints.
  • Path Tracking & Velocity Control: Developed a Pure Pursuit controller integrated with Cross-Track Error (CTE) compensation to ensure high-fidelity waypoint following and stable velocity management.
  • CAN Bus Integration: Utilized ros2_socketcan to interface with the vehicle's control bus. I performed frame decoding using proprietary DBC files to map raw bus signals, including throttle, brake, gear, and steering.
  • By-Wire Controller: Developed a controller node to bridge high-level ROS 2 Twist commands with low-level CAN frames. This included implementing a Finite State Machine (FSM) and ADRC controller to translate target velocities and steering angles into precise actuator setpoints.
  • Environment Mirroring: Engineered a digital twin of the NiFT shuttle and Mcity depot within Gazebo and Rviz2, ensuring accurate physical modeling and sensor data replication.
  • Parity Testing: Structured the simulation nodes to be "plug-and-play" with the physical hardware nodes. This meant the exact same Planner and CAN nodes were used in both sim and reality, significantly reducing deployment bugs.

RViz2 Telemetry: Adaptive Pure Pursuit tracking the generated Bezier trajectory.

Sense
Perception & Localization
LiDAR Point Cloud GNSS / RTK TF2 Transforms
Think
Navigation & Planning
Lanelet2 HD Maps Pure Pursuit Bezier Routing
Act
Communication & Actuation
CAN Bus By-Wire FSM ADRC Controller
Sim
Simulation
Gazebo Digital Twin RViz2 Plug-and-play Nodes

Challenges

CAN Bus & Hardware

Coming into the project with no prior CAN experience, the learning curve for low-level vehicle communication was steep.

  • CAN Frame: I moved beyond high-level code to understand the bit-level details of the DBC file. I had to learn how to manually manipulate bytes to input meaningful velocity and steering data into CAN frames.
  • Linux Networking: A major hurdle was configuring the Linux Kernel to recognize our Kvaser interface. I spent significant time troubleshooting the bridge between can-utils and the hardware, eventually successfully masking our laptop as a native node on the vehicle’s bus.

Software & API

The L4 aspect of NiFT shuttle relies on external data, which introduced unique software integration challenges.

  • Real-time External Comms: I had to implement a Socket.io client within our ROS 2 stack to communicate with the Mcity Octane API. Ensuring that infrastructure-based routing data reached our planner with minimal latency was critical for safe operation.
  • ROS 2 Ecosystem: Beyond just writing nodes, I had to master managing complex Parameter files, launch configurations, and package dependencies to ensure the system was modular and reproducible for the rest of the team.

Updated: