Sandhya Saisubramanian
Oregon State University, OR·Computer Science·Early-career
Robots and other autonomous systems decide what to do using simplified models of the world, but the real world is always messier than any model. A delivery robot or self-driving car will eventually meet situations its designers never planned for, and may then act unsafely or need a person to step in.
- Fields
Computer Science
Engineering
- Topics
autonomous systems
robotics
AI safety
- Grants
- 2 known grants, through Aug 2031
What they work on
- Develop model cognizance methods that let autonomous agents proactively identify and repair limitations in their symbolic decision-making models through limited user interaction.
- Design contextual planning methods that determine "what matters when," prioritizing context-appropriate objectives and balancing tradeoffs in multi-objective planning.
- Build resilient execution techniques that use self-monitoring and online replanning to detect and recover from unforeseen execution-time errors with minimal loss of task performance.
- Create inverse reinforcement learning algorithms that infer reward functions from demonstrations under deontic logic constraints, a formal logic of permission and obligation, to align agent norms with designer intent.
- Design run-time monitors and online metareasoning with introspective perception to predict and correct safety violations, tested on a bipedal robot simulator, scaled autonomous race cars and a campus delivery robot fleet.
More about the research
This research asks how autonomous agents can recognize the limits of their own knowledge, stay safe, and recover from surprises with less human help. One project builds an "introspective" style of reasoning so agents notice and fix their own model gaps; the other develops tools to align a learning robot's sense of what is allowed with what designers intended, and to check that alignment formally.
Why it matters
Robots in healthcare, transportation, manufacturing, warehousing and delivery can fail in ways designers never foresaw, as when an autonomous vehicle swerved erratically to localize itself for lane-keeping. Methods that let systems spot their own model limits, verify that behavior matches design intent, and recover from runtime errors could make deployed robots safer while reducing the human supervision they require.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Jaime Fernández Fisac
Princeton University, NJ·Engineering·Early-career
Develop a framework based on dynamic game theory in which a robot guarantees safety against the worst case within an Operational Design Domain (ODD), the conditions designers or regulators say it must handle, while modeling how agents gain information over time.
Zejian Zhou
University of Wyoming, WY·Engineering
Model complex autonomous systems as multi-agent systems in which a disrupted component becomes an irrational agent, either passive (an accidental fault) or active (adversarial, such as a cyber-attack).
Ioannis Kantaros
Washington University in St Louis, MO·Engineering·Early-career
Train multi-agent reinforcement learning (MARL) control policies for robot teams and mathematically certify their robustness against adversarial, high-dimensional sensor inputs, testing them on aerial and ground robot testbeds and photorealistic simulators.
The school
Oregon State University: 29 labs on PI Pages, $372M a year on research, 238 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.