Interview
Vijay Kumar: Flying Robots | Lex Fridman Podcast #37
- Vijay Kumar, Dean of Engineering at the University of Pennsylvania and former director of the GRASP Lab (established 1979), specializes in multi-robot systems, swarms, and agile autonomous aerial vehicles.
- His graduate thesis involved coordinating a 7,000-pound hydraulic hexapod with 18 joints, each controlled by an independent Intel 8085 processor and 886 coprocessor.
- The first multiprocessor operating system for this hexapod was developed in the 1980s, running on computers with clock speeds of approximately 0.5 MHz.
- Kumar considers the 2011 demonstration of small UAVs forming and deforming three-dimensional patterns in mid-air as his most elegant engineering achievement.
- He distinguishes "aerial robots" from "drones," viewing the latter as a pejorative term implying dumb, pre-programmed devices, whereas the former implies agility and autonomy.
- Ants serve as a primary biological inspiration due to their individual robustness (surviving leg loss) and emergent swarm intelligence regarding colony preservation and consensus.
- Engineering swarms requires a shift from modeling individual high-dimensional components to creating low-dimensional abstractions of group behavior.
- Unlike natural swarms which rely on local, implicit interactions, engineered swarms for specific missions (e.g., surrounding a building) require a global coordinate system and explicit mapping.
- Military and commercial aircraft typically rely on GPS and communication infrastructure, which Kumar identifies as brittle; true autonomy requires navigation without position knowledge, maps, or external pilots.
- The boom in small UAV development post-2007 was driven by the commoditization of Inertial Measurement Units (IMUs), a technology accelerated by federal funding and the automotive mandate for airbag accelerometers.
- Quadcopters utilize an optimal configuration of four motors to control six degrees of freedom by varying motor RPMs to generate thrust and reaction moments for orientation.
- Trajectory planning for autonomous flight must balance safety, optimality, and computational speed to generate smooth motions within time budgets.
- Aerodynamic micro-effects, such as blade flapping and ground/ceiling/wall effects, are difficult to model physically, making data-driven iterative learning necessary for high-speed agility.
- Kumar argues that while machine learning has succeeded in computer vision (perception), fully end-to-end learning from pixels to motor currents for action remains a "holy grail" yet to be fielded.
- He warns that scaling computer vision accuracy to safety-critical levels (e.g., 99.9%) requires an exponentially increasing amount of data, leading to unsustainable energy consumption from data farms (2% of US electricity in 2014).
- Kumar asserts that autonomous flight is mathematically simpler than autonomous driving because flight allows for vertical trajectories to avoid 2D obstacles, despite the complexity of 3D aerodynamics.
- Widespread delivery drone adoption is currently limited by battery technology rather than autonomy, specifically regarding the high power density required for vertical takeoff (approx. 200 watts/kg vs. 80 watts for a human brain).
- Regarding "flying cars," Kumar predicts electric vertical takeoff and landing (eVTOL) vehicles will be viable for short distances, but fossil-fuel-powered options will likely be more economically viable initially.
- Human-robot collaboration is essential, with distinct modes including human command, shared autonomy (collaboration), and human as a bystander (e.g., in search and rescue).
- Kumar critiques the Tesla "Level 2" approach where humans monitor the machine, citing human boredom as a major failure point, and prefers shared autonomy for unstructured, high-stakes tasks.
- A significant research opportunity involves "AI-enabled psychology," where robots model human state (attention, fatigue) to enhance safety rather than just modeling the environment.
- Kumar expresses concern over the weaponization of swarms, noting that defensive capability against adversarial swarms is as critical as the offensive technology itself.
- He advocates for "technology literacy" as a new liberal art, urging that engineers and politicians must understand the societal implications of autonomous systems.
- The most significant open problem in robotics is generalizing systems to operate in unstructured, variable environments (e.g., Mumbai or Naples) rather than structured settings like parking lots.
- Kumar advises engineering undergraduates to prioritize mathematical foundations and representation over existing software packages, and to cultivate breadth in the liberal arts.
- He emphasizes that while board games are structured environments where AI succeeds, the physical world requires handling unstructured uncertainty that cannot be solved by simply finding a deep neural network.