newsfilter.io
Lecture, Conference Presentation

Sertac Karaman (MIT) on Motion Planning in a Complex World - MIT Self-Driving Cars

  • Speaker Profile: Sirtesh Karaman, MIT AeroAstro professor specializing in autonomous vehicles (land/air, 2–18 wheels) using formal optimization and deep learning; programs on Kinesis keyboards with Emacs.
  • Historical Context: The DARPA Urban Challenge (2007) marked a decade-long transition from academic university projects to industry-changing technology; MIT was one of six finishers out of 89 teams.
  • Algorithmic Evolution:
    • The original Rapidly Exploring Random Tree (RRT) algorithm was found to fail to converge to optimal solutions in constrained environments, often trapping robots in local loops.
    • Karaman's doctoral thesis (2011) introduced RRT*, an asymptotically optimal variant that guarantees convergence to the best trajectory with negligible computational overhead compared to standard RRT.
    • RRT* naturally generates complex maneuvers (e.g., skid-to-turn race car dynamics) without explicit programming.
  • System-Level Optimization:
    • Autonomous vehicle systems (e.g., Amazon Kiva warehouses) operate efficiently only below a "critical density"; above this threshold, systems become fragile and prone to cascading collisions.
    • High-dimensional control strategies are compressed from petabyte-scale lookup tables to ~2 MB using singular value decomposition and data compression techniques.
    • Future logistics could reduce transport costs in dense cities to ~$0.50–$0.99 per trip via shared, autonomous, electric mobility-on-demand networks.
  • DARPA Urban Challenge Technical Implementation:
    • MIT's entry utilized a Land Rover LR3 with 5 cameras, 16 radars, 12 planar laser scanners, 1 3D Velodyne scanner (64 lasers), and a 40-CPU server rack with a mobile generator and AC unit.
    • The software stack consisted of ~200,000 lines of C code across ~100 processes; lightweight communications and marshaling (LCM), built for the challenge, is now industry standard (Ford, Toyota).
    • A collision with a stuck Cornell vehicle occurred due to a map refresh logic error: the collision-checking algorithm assumed static obstacles and failed to update the map when no new sensor data arrived.
  • Current Research Group Focus:
    • Vehicle Level: Developing ultra-agile robots (e.g., 200Hz camera drones, falcon-mimicking interceptors) requiring kilohertz real-time control and high-performance onboard GPUs.
    • Perception: Utilizing visual odometry with as few as 10 features for computationally efficient localization; testing autonomous tricycles in Taiwan for deep learning data collection.
    • System Integration: Projects include autonomous golf carts in Singapore, autonomous wheelchairs, and MIT-Singapore electric vehicle integration.
  • Industry Trends & Challenges:
    • Academic collaboration has declined due to commercial competition and secrecy, though MIT researchers continue publishing on failure modes (e.g., joint paper with Cornell).
    • Transition from sensor suites (LIDAR-heavy) to vision-only (camera-based) navigation is expected to mature within 3–5 years, though deep learning remains the dominant approach.
    • Autonomous driving faces non-technical barriers including infrastructure architecture, insurance regulations, cybersecurity (for V2V/V2I communication), and ethical policy considerations.
  • Entrepreneurial Activity: Karaman is a PI on Optimus Ride (seeded at >$5M), a company aiming to deploy autonomous vehicles for on-demand urban transport, featuring a founding team with backgrounds in urban planning, design, and former Google X leadership.
  • Forward-Looking Statements:
    • High-speed/low-complexity environments (highways) are projected to be solved within 3 years.
    • Low-speed/high-complexity environments (parks, campuses) offer quicker opportunities for specific autonomous applications.
    • Vision-only autonomy without LIDAR or mapping is predicted to be viable in 10 years, with integrated model-based and deep learning approaches likely dominating the interim period.
    • Vehicle-to-Infrastructure (V2I) and Vehicle-to-Vehicle (V2V) communication could enable lane-following within 3–5 years, though intersection autonomy remains further out.