Lecture, Conference Presentation
Sertac Karaman (MIT) on Motion Planning in a Complex World - MIT Self-Driving Cars
- The transformation of academic DARPA Urban Challenge initiatives into a world-changing industry is expected to require approximately 10 years, with highway autonomy in low-complexity environments targeted for realization within the next three years.
- Motion planning capabilities are projected to advance such that fully autonomous driving using only cameras without LiDAR or pre-mapping could be achieved in three to five years or within a decade, though not within the next year.
- Vehicle-to-vehicle (V2V) and vehicle-to-infrastructure communication features, such as lane following, are predicted to become viable within three to five years, pending the resolution of cybersecurity barriers.
- Autonomous systems operating above a "critical density" threshold face risks of fragility and failure from single component crashes, indicating that current warehouse and port multi-vehicle setups are not yet robust enough to handle these scenarios.
- Computational constraints for six-dimensional drone controllers, which could naively require 2.5 petabytes, can be reduced to approximately two megabytes using singular value decomposition, allowing supercomputer-computed controllers from five-minute runs to be deployed as lookup tables for kilohertz execution.
- The evolution of the RRT algorithm to RRT* is expected to provide asymptotic optimality and enable complex maneuvers like high-speed vehicle skidding, addressing fundamental flaws in the original algorithm where convergence to optimal solutions often failed.
- Transportation economics are forecast to shift toward 99 cents per trip with a five-minute wait time, 50 cents with ride-sharing, and 30 cents for single stops, driven by increased efficiency and autonomy.
- Development efforts currently include five autonomous tricycles in Taiwan with a target of expanding to 30 units for deep learning data collection, alongside projects integrating autonomous forklifts, electric vehicles, and wheelchairs.
- Optimus Ride, a co-founded entity, has secured slightly more than $5 million in seed funding to initiate operations, while the speaker also advises a Formula SAE team utilizing deep learning for autonomous racing.
- While complex environments like parks or university campuses remain distant from full autonomy, specific applications in these areas are identified as holding future opportunities.