Research Projects

Artificial Intelligence enabled autonomous systems are becoming increasingly capable. In the Indian context, AI-based robots will be deployed alongside humans in an augmented intelligence modality. The primary barrier to their widespread deployment is the lack of safety guarantees. The national strategy for AI has identified tremendous potential in deploying human-autonomy systems in mission-critical situations such as disaster response, industry 4.0, and defense. However, given the current lack of safety guarantees, human operators are having to spend costly person-hours in: 1) programming safe behaviours, 2) acting as backup safety operator, and 3) direct teleoperation. Safe autonomous systems will allow human operators to perform more tactical and strategic decision making. While algorithms for safe operation grounded in solid theoretical foundations exist, they have largely been shown in simulation. Crucially, existing approaches assume complete state information from rich sensors, which is unmet in the real world. There is a research gap in safe policy learning for autonomous systems purely based on visual state representation. We seek to leverage the recent advancements in foundation models such as vision-language models (VLMs) along with safe reinforcement learning algorithms. Scientific objectives: SO1) To elicit safety guidance from VLMs in the form of: 1) Unsafe state identification, 2) Constraints of safe operation, and 3) Safety ranking of proposed actions. SO2) To incorporate VLM safety outputs into safe reinforcement learning algorithms to train safe visuomotor policies for autonomous navigation in real-world environments SO3) To enable autonomous systems to explain safety related decisions governed by the underlying constraints SO4) To demonstrate the system on an autonomous ground vehicle in a real-world navigation scenario Hypotheses: H1) VLMs would be able to provide safety guidance to autonomous agents H2) Safety guidance can be used as constraints to train visuomotor policies through safe reinforcement learning algorithms H3) Autonomous systems will explain their behavior, leading to better integration and trust with humans Main experiments: E1) Iterative prompting strategies to elicit safety information from vision language models E2) Testing visuomotor policies in simulation on standard safe reinforcement learning benchmarks E3) Demonstration on an autonomous ground vehicle in a real-world navigation scenario Significance to the field: If the objectives are met, the approach will be applicable to a variety of applications such as disaster response, industry 4.0, and defense. Further, the successful execution of the project will pave the way for a paradigm shift towards safe visuomotor policies based on foundation model guidance. This will lead towards reliable autonomous operation in challenging real-world scenarios and enable human operators to take on more tactical and strategic decision-making roles.