Language model evaluation
2 papers
Independent work on ranking language models and testing their advice.
Preprint
Bakshi, S., Bose, A., Bose, A., and Nandi, S. · arXiv preprint, 2026
People often give a different answer to the same risk-tolerance question when asked again, and seven LLM advisers shift a client's stock allocation by 24 to 35 points when only that answer changes. Telling the models the answer is noisy has side effects, while a single clarifying question removes nearly all of the swing.
Both of these started when Apurba took Stanford’s CS329H, Machine Learning from Human Preferences.
Preprint
Bakshi, S., Bose, A., Bose, A., and Nandi, S. · arXiv preprint, 2026
Across about 1.8 million Chatbot Arena votes, matchups between versions of the same model are more decisive than a single-score ranking predicts, tracking how alike the two models' answers read. A small correction predicts held-out votes better, mostly through ties, while the leaderboard order barely moves.
Robot fleets: scheduling and planning
11 papers
My PhD work and its follow-ons: getting teams of autonomous mobile robots to the right tasks, in the right order, using as little time and energy as possible.
Conference
Li, J., Bakshi, S., Chu, J., Li, S., and Chen, D. · IEEE International Conference on Systems, Man, and Cybernetics (SMC), 2026
A two-stage approach for robot fleets: a sequential auction with closed-form bids hands out tasks, then each robot plans an energy-minimal trajectory using a physics-based battery model, with collision refinement and event-triggered rescheduling.
Conference
Chu, J., Huang, J., Bakshi, S., Zhu, Y., Ohman, E., and Chen, D. · American Control Conference (ACC), 2024
Plans robot trajectories through factories and warehouses with energy use as the objective, while respecting the robot's configuration constraints.
Conference
Chu, J., Zhao, F., Bakshi, S., Yan, Z., and Chen, D. · American Control Conference (ACC), 2021
Brings unpredictable human operators into robot path planning, so robots sharing a floor with people choose routes that account for the risk of interaction.
Journal
Bakshi, S., Feng, T., Yan, Z., Ma, Z., and Chen, D. · Journal of Intelligent & Robotic Systems, 2021
Plans a mobile robot's trajectory and recharging to minimize energy across a task space where travel costs differ by direction.
Conference
Chu, J., Bakshi, S., Qin, H., Yan, Z., and Chen, D. · ASME Dynamic Systems and Control Conference (DSCC), 2020
A low-cost hardware and software platform for developing and testing control and routing algorithms on automatic ground vehicles.
Thesis
Bakshi, S. · PhD dissertation, UT Austin, 2020
Ties the work together: continuously assigning prioritized tasks to a fleet of robots, scheduling them in real time with convex relaxations and reinforcement learning, and planning energy-optimal trajectories.
Conference
Bakshi, S., Feng, T., Yan, Z., and Chen, D. · American Control Conference (ACC), 2019
Relaxes the combinatorial single-robot scheduling problem into a regularized quadratic program, so prioritized tasks can be ordered fast enough for real-time use.
Written on a flight back to India. I was bored and had forgotten to download any movies.
Conference
Woo, S., Keller, A. M., Choudhry, W., Trakhtengerts, R., Bakshi, S., Yan, Z., and Chen, D. · American Control Conference (ACC), 2019
A test platform for routing a ground vehicle in real time as tasks with different priorities arrive at random.
Conference
Bakshi, S., Feng, T., Yan, Z., and Chen, D. · ASME Dynamic Systems and Control Conference (DSCC), 2019
A probabilistic method for deciding, in real time, which robot in a fleet should take each prioritized task.
Journal
Bakshi, S., Feng, T., Yan, Z., and Chen, D. · ASME Journal of Dynamic Systems, Measurement, and Control, 2019
Replaces slow exhaustive search with a fast scheduler that lets a large fleet of robots finish prioritized tasks in near-minimum time.
Journal
Bakshi, S., Yan, Z., Chen, D., Qian, Q., and Chen, Y. · ASME Journal of Dynamic Systems, Measurement, and Control, 2018
A two-step algorithm that schedules a factory ground vehicle's visits as a traveling-salesman problem, much faster than exhaustive search.
My first paper from my PhD work. It came out of a course project.
Data-driven and learning-based control
5 papers
Recent work with UT Austin collaborators on controllers built from data, and on tracing a controller's behavior back to the data that shaped it.
Journal
Li, J., Li, S., Bakshi, S., Xu, J., and Chen, D. · ASME Letters in Dynamic Systems and Control, 2026
Borrows influence functions from machine learning to show which identification trajectories help or hurt an LQR controller designed from the identified model.
Accepted
Li, J., Li, S., Bakshi, S., Xu, J., and Chen, D. · IEEE Conference on Decision and Control (CDC), 2026
Extends trajectory-level data valuation to stochastic LQR, accounting for how each trajectory shifts both the learned dynamics and the estimated noise.
Preprint
Li, J., Li, S., Bakshi, S., Xu, J., and Chen, D. · arXiv preprint, 2026
Traces failures of behavior-cloned neural controllers back to the training data, with influence variants adapted to closed-loop dynamics and safety constraints.
Preprint
Li, J., Duan, X., Li, S., Bakshi, S., and Chen, D. · arXiv preprint, 2025
Learns to predict how much each sampled trajectory matters in MPPI control, so the controller can focus on useful samples in real time.
Conference
Li, S., Li, J., Martin, C., Bakshi, S., and Chen, D. · IFAC-PapersOnLine (Modeling, Estimation and Control Conference), 2025
Predictive control for systems that are only partly known: model constraints cover the known dynamics, data covers the rest, and distributionally robust optimization handles uncertainty.
Manufacturing process control
3 papers
Control for a roll-to-roll process that could make graphene and flexible electronics at industrial throughput.
Journal
Martin, C., Zhao, Q., Bakshi, S., Isaza Velásquez, E., Li, W., and Chen, D. · ASME Journal of Dynamic Systems, Measurement, and Control, 2024
Wraps the nonlinear peeling dynamics in a tighter convex set (a norm-bounded linear differential inclusion) to design an optimal controller for the dry-transfer process.
Conference
Martin, C., Zhao, Q., Bakshi, S., Chen, D., and Li, W. · IFAC-PapersOnLine, 2022
Holds the peeling angle steady with an H∞ controller, a key requirement for making roll-to-roll dry transfer practical.
Conference
Martin, C., Zhao, Q., Bakshi, S., Chen, D., and Li, W. · IFAC-PapersOnLine, 2021
Introduces a polytopic convexification of the peeling-front dynamics and shows it supports better control than a plain linearization.
Biomedical systems
4 papers
Dynamic-systems tools applied to the heart: predicting bradycardia in preterm infants, controlling a blood pump, and cleaning up digitized ECGs.
Accepted
Li, J., Li, S., Bakshi, S., Li, W., and Chen, D. · IEEE International Conference on Bioinformatics and Biomedicine (BIBM), 2026
Wraps any ECG digitizer with closed-loop vision-language-model feedback, grounded in quantitative signal checks, to catch failures on real-world scans.
Accepted
ECG-OOD-Trust: Auditing Cross-Dataset Generalization, Explanation Faithfulness, and Shortcut Leakage in 12-Lead ECG Models
Li, J., Li, S., Bakshi, S., Li, W., and Chen, D. · IEEE International Conference on Bioinformatics and Biomedicine (BIBM), 2026
Audits 12-lead ECG models for how well they hold up on new datasets, whether their explanations reflect what drives their predictions, and whether they lean on shortcuts.
Journal
Bakshi, S., Feng, T., Chen, D., and Li, W. · ASME Journal of Engineering and Science in Medical Diagnostics and Therapy, 2020
Treats the infant cardiovascular system as a dynamic system and uses system identification to predict episodes of slowed heart rate before they happen.
Thesis
Bakshi, S. · MS thesis, UT Austin, 2017
A hybrid model, observer, and robust sliding-mode controller for piston transitions in the TORVAD ventricular assist device.
Drilling dynamics
3 papers
Modeling and control for directional drilling in oil and gas, from my early PhD years.
Journal
Gu, Q., Fallah, A., Feng, T., Bakshi, S., Chen, D., et al. · Journal of Petroleum Science and Engineering, 2020
A control strategy for handling gas kicks in deepwater wells and reducing the risk of gas unloading in the riser.
Journal
Feng, T., Bakshi, S., Gu, Q., Yan, Z., and Chen, D. · Journal of Petroleum Science and Engineering, 2019
Optimizes the design of the drill string's bottom-hole assembly to reduce harmful vibration while drilling.
Journal
Feng, T., Bakshi, S., Gu, Q., and Chen, D. · ASME Journal of Computational and Nonlinear Dynamics, 2019
A finite element framework for curved beams under heavy compression and wall contact, motivated by directional drilling.
Vehicle dynamics and powertrain
4 papers
Undergraduate work on IIT Bombay's Formula Student electric race car.
Conference
Makhija, G., Mahajan, M., Bansod, N., and Bakshi, S. · ASME IDETC/CIE, 2015
Optimizes the powertrain of a rear-wheel-drive electric race car around a hybrid planetary gearbox to cut weight without losing performance.
Conference
Sakhalkar, S. V., Dhillon, P., Kumar, P., Bakshi, S., and Arora, P. S. · SAE Technical Paper, 2014
A closed-loop electronic differential that vectors torque between the rear wheels using steering, throttle, wheel-speed, and yaw feedback to improve stability and response.
Journal
Bakshi, S., Dhillon, P., and Maruvada, T. · SAE International Journal of Materials and Manufacturing, 2014
An iterative method for picking a light, compact, high-speed planetary gearbox configuration for a Formula Student car.
Conference
Sakhalkar, S. V., Dhillon, P., Bakshi, S., Kumar, P., and Arora, P. S. · SAE Technical Paper, 2014
A powertrain model used to choose the gear reduction ratio and estimate the battery energy an electric race car needs.