Robotics · Reinforcement Learning

Abhishek S

The hard part of robot learning is not the learning. It is the data.

Research Intern, Indian Institute of Science
with Prof. Pavankumar Tallapragada

I spend most of my time on robots: mapping them, coordinating them, benchmarking them, and lately collecting the data they will learn from. Five first-authored workshop papers at ICRA and RSS 2026 and a startup came out of that.

◆ Applying for MS / PhD, Fall 2027

Email GitHub Scholar LinkedIn CV

01News

Jul 2026EA-LNS ("Beyond Sum-of-Costs") accepted as an oral at the SBO Workshop, RSS 2026. Fifth first-authored workshop paper of the cycle.
Jun 2026Subterra accepted at the WPCIS Workshop, RSS 2026. Fourth first-authored workshop paper of the cycle.
Jun 2026Preprints of SENTINEL, CADENCE, and MineXplore posted to arXiv.
May 2026SENTINEL, CADENCE, and MineXplore accepted at ICRA 2026 workshops (UOWR, MARS-EAI, and Xplore).
Jan 2026Founded a robotics startup, building in stealth: curated real-world robotics data for foundation-model post-training.
Aug 2025Joined Prof. Pavankumar Tallapragada's group at IISc, working on real-time multi-robot coordination.

02Publications

Five first-authored workshop papers across ICRA 2026 & RSS 2026.

ICRA 2026

SENTINEL: Teaching Robots to Say “I Don't Know” — Uncertainty-Aware SLAM

Abhishek S, Badrikanath Praharaj, Sreeram MV

Uncertainty in Open-World Robotics (UOWR) Workshop, ICRA 2026 · Accepted

arXiv

ICRA 2026

CADENCE: Predicting Realized MAPF Execution Time Beyond Sum of Costs

Abhishek S, Badrikanath Praharaj, Sreeram MV

Multi-Agent Robotic Systems: Real-World Collaboration & Interaction (MARS-EAI) Workshop, ICRA 2026 · Accepted

arXiv

ICRA 2026

MineXplore: An Open-Source RL Exploration Benchmark for GNSS-Denied Underground Environments

Abhishek S, Badrikanath Praharaj, Sreeram MV

Xplore Workshop (Exploration in Robotics, RL & Search), ICRA 2026 · Accepted

arXiv

RSS 2026

Subterra: A Validated Benchmark for GNSS-Denied Tunnel Navigation

Abhishek S, Badrikanath Praharaj, Sreeram MV

WPCIS Workshop, RSS 2026 · Accepted

arXiv soon

RSS 2026

Beyond Sum-of-Costs: Plan-Space Sampling for Execution-Time Optimization in MAPF

Abhishek S, Badrikanath Praharaj, Lakshmi Prasad Seetharamu, Sreeram MV

Sampling-Based Optimization (SBO) Workshop, RSS 2026 · Accepted (Oral)

arXiv soon

Where this is going → full versions of MineXplore and Subterra are being extended toward ICRA 2027; CADENCE toward IEEE RA-L; and SENTINEL toward IEEE T-RO.

03Experience

Indian Institute of Science (IISc)Aug 2025 – Present

Control & Optimization InternBengaluru, India

Designed a real-time multi-robot coordinator on a custom quadrant-token abstraction that partitions shared grid cells into sub-cell waypoints, robust to spatio-temporal conflicts and collision-free under quadrant mutual exclusion. Validated across 100+ live OptiTrack experiments (4 robots, scalable to 10), cutting average inter-robot wait to ~0.1 s. Advised by Prof. Pavankumar Tallapragada.

Code ↗

4 robots — live OptiTrack run

7 robots — coordinated paths

Stealth Robotics StartupJan 2026 – Present

FounderBangalore, India

Building in stealth: an early-stage robotics-research venture tackling the scarcity of real-world robotics data for foundation-model post-training, through curated datasets for niche, hard-to-source domains that established pipelines underserve. More soon.

04Projects

01

EdgeLIOPES University, Bangalore

SLAM / Perception

Onboard LiDAR–inertial odometry and mapping (FAST-LIO2, Livox MID360, source-built ROS 2 Humble) running entirely on a Jetson Xavier NX — no offboard compute. Real-time state estimation and 3D reconstruction: 10 Hz mapping sustained over a 15-min indoor run at ~22% CPU, 15 ms/scan latency, and 0.13 m ATE RMSE.

GitHub

02

Event-Triggered Lyapunov-Safe Reinforcement LearningIISc, Bangalore

Research Project · Safe RL / Control

Cast event-triggered control of a linear system as a constrained MDP; learned safe, state-dependent triggering via Q-learning → DQN → SAC under a Lyapunov certificate with certified-set projection (Zeno-free). Proved the optimal inter-event time has intrinsic jump discontinuities in the trigger parameter β. Advised by Dr. P. Tallapragada.

03

SceneForgePES University, Bangalore

Interactive 3D / Digital-Twin Tooling

In-browser 3D transform editor (React Three Fiber, Three.js): draggable numeric inputs for position, rotation, and scale in real-time two-way sync with an in-viewport gizmo, with per-axis increment locking and a movable properties panel.

GitHub Live demo

EdgeLIO◱ hover to preview

EdgeLIO

05Education

PES UniversityDec 2022 – Present

B.Tech, Computer Science & EngineeringBangalore, India · GPA 7.7 / 10