Shipping autonomous systems from simulation to the real world - production fleet deployment (200+ AMRs), real-vehicle autonomy, and sim-to-real RL for humanoid robots. MEng in Autonomy & Robotics at UIUC.
sunnydeshpande9900@gmail.com Conversational navigation on a Unitree Go1 with Jetson AGX Orin 64GB. Frozen NaVILA (8B VLA) consumes 8 ZED RGB frames and emits mid-level language actions (e.g. "move forward 75cm") at ~0.1 Hz, parsed to Twist cmd_vel.
A token-level confidence layer flags uncertain or degenerate outputs and triggers natural-language clarification or safe fallback. Offboard inference bridged via rosbridge WebSocket over Tailscale.
Two-policy USB-A insertion on a Unitree UR5e. OpenVLA handles the coarse approach via Cartesian deltas through MoveIt Servo, while a SAC residual policy learns fine corrective joint deltas for tight-tolerance insertion.
A force-hysteresis contact gate (2N on, 1N off) activates the residual only on contact. Trained in MuJoCo with domain randomization, deployed via ROS2 Humble.
Hierarchical PPO locomotion policy for a 29-DOF Unitree G1 humanoid, trained in Isaac Lab with a 7-phase terrain curriculum. Uses teacher-student distillation to compress a privileged expert policy into a deployable student that relies only on onboard height scan data.
Active pedestrian avoidance system on the Polaris GEM e4, building on an existing LiDAR and vision perception stack. Implements an MPPI trajectory planner with a diffusion-based pedestrian prediction module for continuous, intent-aware avoidance.
Real-world deployments, novel algorithms, and sim-to-real systems across humanoids, AVs, and mobile robots. Featured projects showcase the deepest work.
Replaced separate EfficientNet + BERT encoders with a frozen unified CLIP backbone (128-dim adapter projection) in the HumanVLA teacher–student pipeline, resolving catastrophic forgetting observed when fine-tuning visual encoders. Behavior Cloning + DAgger distillation across 615 episodes in 4 HITR room types; 62.1% success on unseen tasks (+1.9% over baseline, −4.4% placement error).
Modular ROS 2 autonomy stack on the UIUC Polaris GEM e4: Ouster OS1-128 LiDAR → DBSCAN clustering + EMA tracking fused with OAK-D LR RGB-D via YOLOv11 detection (weighted 0.8/0.2 distance, 0.3/0.7 bearing). Forward-simulated TTC drives a 3-state safety FSM; Stanley lateral + PID longitudinal control with PACMod2 hard-brake override. 91% field success.
Two-tier PPO hierarchy for dual Unitree G1 humanoids in IsaacLab: low-level [256,256] MLP drives 29-DOF joint-position-delta actions through a 7-phase standing→omnidirectional curriculum (30M steps, 256 envs); high-level [64,64] navigation policy converges agents from 10 m to 0.5 m with 95% success. 25+ shaped reward terms enforce gait symmetry, contact penalties, and energy regularization.
Open-vocabulary 6D tracking extending FoundationPose with Moondream2 VLM scene analysis, SAM-3 text-prompted segmentation, and hierarchical mesh acquisition (GT → Objaverse-XL retrieval → TripoSR generation). Composite scoring (IoU + Depth + Silhouette) selects best proxy; achieves 100% ADD-S AUC on texture-rich objects and supports language-driven dynamic target switching mid-task on YCB-Video.
Full-stack embodied AI on physical Booster K1 Humanoid: Gemini VLM semantic goal planning with ROS2 Nav2 autonomous navigation and Snap AR Spectacles AR collaboration.
Led 7-person team. Kalman-filtered admittance controller maintaining <150 N brush-to-glass contact. Full ROS2 stack on Jetson Orin Nano with SMACH FSM. Validated in live Changi Airport deployment.
Novel PP variant incorporating bicycle-model lateral error decomposition (cross-track slippage + heading error) into the control law. DIAPP adds a dynamics correction term c_d·e_lat,total to the steering output; RDIAPP extends this with RPP-style curvature-regulated velocity. At 70 km/h on an S-track, RDIAPP achieves 1.79 m max CTE vs. PP's 30.26 m (diverges); on a U-track, 2.49 m RMS CTE vs. 31.6 m for PP.
Autonomous lane-following on Jetson Xavier + ZED2: dual-filter perception (Gaussian-weighted adaptive grayscale + HLS color space) with bitwise AND fusion targeting 8% pixel density, vanishing-point radial-scan histogram rejecting racetrack markings, and sliding-window 2nd-order polynomial fit. Custom proportionate lateral controller with distance-weighted steering (f(c_y) = C/(c_y − h_y + ε)). Completed full 12-min autonomous laps across variable lighting conditions.
Autonomous unknown-maze exploration on TurtleBot3 (LattePanda): frontier-based exploration with DFS cluster grouping and centroid-based goal selection, A* costmap planner with B-spline path smoothing, Pure Pursuit trajectory follower, and fine-tuned YOLOv11n with a novel spatial deduplication tracker projecting bounding-box centroids into the map frame via LiDAR range lookup for unique instance counting.
Software Lead. Cascaded PID for 6-DOF underwater stability with IMU/depth fusion via state-space control. Real-time CV for autonomous target acquisition. SAUVC 2023 Finalist.
Systematic evaluation of 7 monocular and stereo VO algorithms (DSO, SVO, CNN-SVO, DF-VO, TartanVO, ORB-SLAM3, DROID-SLAM + our heuristic variant) across 3 geographically-diverse rain datasets (Oxford RobotCar, 4Seasons Munich, Singapore heavy rain) using Absolute Trajectory Error. DF-VO performed best for monocular <500 m; our DROID-SLAM + CGRP + Heuristic approach achieved the lowest stereo ATE for long-range localization. Published at IEEE CASE 2023.
CPG-based motion controller for soft batoid robot with sinusoidal propulsion and strain gauge obstacle detection.
CNN-RNN model predicting Cat 1 Lightning Risk from weather station features: rainfall, wind speed, temperature, humidity, wind direction.
Contextual transformer embeddings + sequential modeling. ~96% accuracy outperforming classical NLP baselines.
Hidden Markov Models with first- and second-order Viterbi decoding for entity recognition and sentiment tagging.
Extracted and benchmarked 9 nonlinear feature families (MMSE, RQA, DFA, entropies, FD, Hjorth, Hurst, LLE, LZC) from 7,500 focal/non-focal EEG signals (Bern-Barcelona database). LS-SVM (polynomial-3) with 10-fold CV achieved 87.93% accuracy, 89.97% sensitivity. MMSE ranked highest by t-test significance. Published in FGCS, Elsevier.
Low-cost electromagnetic leak detector with Bluetooth connectivity using EM conduction principles for non-invasive detection.
Custom PCB with motor control, dynamic air-cooling, slope-based actuation, and anti-roll basket mechanism.
Low-cost quadcopter with off-the-shelf components. Integrated propulsion, power distribution, and flight electronics for stable teleoperation.
Angsana seed-inspired autorotating craft with single-flap motor trajectory control and aerodynamic calibration.
8-bar lifting mechanism with novel intake for cube frame collection. Won Best Design Award at Asia-Pacific Robotics Championship.
Production deployments, real-vehicle autonomy, and simulation at scale - from 200+ AMR fleets to port-side AV trials.
Full-stack robotics - from perception and planning to learning and deployment.
Spoken Languages: English (Fluent) · Hindi (Fluent) · Marathi (Fluent)
A focused academic path through engineering, robotics, and CS.
Peer-reviewed contributions to robotics and biomedical AI.
Comprehensive evaluation of 7 VO algorithms (DSO, SVO, CNN-SVO, DF-VO, TartanVO, ORB-SLAM3, DROID-SLAM) on monocular and stereo setups across Oxford RobotCar, 4Seasons (Munich), and Singapore heavy-rain datasets. Proposed DROID-SLAM + CGRP + Heuristic variant achieving lowest stereo ATE for long-range rain localization; DF-VO identified as best monocular approach for <500 m.
Download PaperFirst systematic comparison of 9 nonlinear feature families (52 features, all p < 0.01) for focal vs. non-focal EEG classification on the full 7,500-signal Bern-Barcelona database. LS-SVM (polynomial-3, 10-fold CV) achieved 87.93% accuracy / 89.97% sensitivity. MMSE identified as top-ranked discriminator; proposed recurrence, bispectrum, and cumulant plots for visual class separation. DOI: 10.1016/j.future.2018.08.044
Download PaperTailored resumes highlighting relevant experience for each role.
My passion to learn about how the world and the things around me work has been present since childhood. I've always been fascinated by machines and how they work. This has driven me to choose robotics and artificial intelligence as my fields of study. Ever since, I've been enjoying making things come to life.
I'm currently a Master's student at University of Illinois Urbana-Champaign studying Autonomy and Robotics. I hold a BEng from SUTD (Singapore) with a Robotics Focus and CS Minor under the Honours and Research Programme (SHARP) with Global Distinguished Scholarship. My industry experience spans production fleet deployment of 200+ AMRs at Hyundai, AV simulation at Venti Technologies, and perception research at A*STAR. I'm driven by shipping systems that work in the real world.
Fluent in English, Hindi, and Marathi.
Building robots (of course), tinkering with electronics, photography, exploring new places and cultures, martial arts, and staying curious about how everything works.
