Aditya Kothari

Robotics & Autonomy Researcher

Making robots learn skills from vision alone

Stanford University Multi-Robot Systems Lab  ·  MS AeroAstro  ·  GPA 4.0
View Research Resume GitHub
Aditya Kothari
Stanford MSL Research Agentic policy training · 3D Gaussian Splats
Selected Work Visuomotor policies · RL · state estimation
Archer Aviation Multi-agent planner for swarm-level search

01. Research

I'm an MS student in Aeronautics & Astronautics at Stanford, working in the Multi-Robot Systems Lab under Prof. Mac Schwager. I build visuomotor policies for flight: systems that map raw onboard camera frames to control, trained in photorealistic 3D Gaussian Splat reconstructions and transferred to hardware. Most recently I was an autonomy intern at Archer Aviation, where I built a multi-agent planner for swarm-level probabilistic search. Before Stanford I graduated top of my class from VIT, led a 25-member university robotics team that fielded 7 UAVs and ground robots, and held engineering roles at Boeing and Tata Motors.

Current Research at Stanford MSL ● Active since Jan 2026

AORA: Closing the Loop on Autonomous Policy Training

Stanford Multi-Robot Systems Lab  ·  Advisor: Prof. Mac Schwager

Leading development of AORA, an agentic orchestrator that turns policy training into a closed self-improvement loop: it provisions cloud compute, trains and evaluates vision-language-action (VLA) policies through an agentic “Forge,” and feeds failures back into the next round without a human supervising, with full sim-to-real onto a Unitree G1 humanoid. Alongside it I design system-2 visuomotor policies for language-guided drone navigation, trained entirely in 3D Gaussian Splat simulation, removing the 60+ hours of manual flying such policies would otherwise require.

Agentic Orchestration Visuomotor Policies 3D Gaussian Splats Language-Guided Navigation Sim-to-Real Transfer

Broader Research Interests

Robot Learning State Estimation Vision-Language Models Autonomous Navigation Multi-Robot Systems Hierarchical Planning Perception for Autonomy

02. Selected Work

V-LEAD overview: a closed-loop rollout to a mannequin goal with the onboard RGB the policy consumed, the RGB-to-velocity architecture, the same goal reached from many start points, and policy-vs-expert trajectories for three targets
Visuomotor Policy · RL Jun 2026

V-LEAD: Vision-Only Drone Navigation in Gaussian Splats

A quadrotor that flies to a goal through cluttered space using nothing but a forward camera. No map, no depth sensor, no full state estimate. Team project with 2 teammates.

  • Distilled a privileged MPC expert into an RGB-only visuomotor policy: ResNet-18 fused with a frozen Depth Anything V2 branch via cross-attention, a GRU for temporal context, and a 10-step velocity-command horizon at 20 Hz
  • Trained by behavior cloning, corrected with DAgger using an MPC re-solve oracle, then pushed past the demonstrator with online RL (SAC, PPO, residual TD3+BC)
  • My ownership: the 3DGS simulator integration, the expert-data pipeline, and the offline-to-online RL stack, including the frozen-backbone residual actor, critic warmup, and BC anchor needed to stop RL from erasing the imitation seed
Key Outcome: Goal success rose from 14–52% to 75–97% across the pipeline while collisions fell 33% → 5%. On a goal object held out of training entirely, residual TD3+BC was the only method to improve on the imitation seed (65.1% vs 62.4%) where on-policy PPO regressed to 40%.
PyTorch3D Gaussian SplattingTD3+BCSACDAggerDocker
View on GitHub →
Tracked pedestrians with persistent IDs in a CARLA urban scene
Multi-Object Tracking · Estimation Mar 2026

MAVBE: Behavior-Aware Multi-Object Tracking

An IMM-based tracker that blends motion models and social dynamics to accurately track multiple objects in 3D.

  • Interacting Multiple Model (IMM) filter blending CV, CT, and CA motion models with Social Force augmentation for behavior-aware trajectory prediction
  • CV/CT/CA EKF updates, camera-to-world transforms, and Markov transition matrix calibration within a modified DeepSORT pipeline with depth unprojection for full 3D tracking
  • Validated on real-world video sequences and CARLA simulation at varying object densities
Key Outcome: 55% RMSE reduction (0.50m vs 1.10m) and 68% fewer identity switches vs. vanilla Kalman baseline at 5-object density.
IMM-EKFDeepSORTYOLOv9CARLAPython
Localization · Space Robotics Mar 2026

StarNav: GPS-Free Lunar Rover Localization

Probabilistic localization for lunar rovers using star pattern recognition and particle filtering from sky imagery alone.

  • Full pipeline: multi-threshold blob detection → coordinate transforms → lunar ENU frames → geometric star triangle signatures for robust star pattern recognition
  • Particle filter localization across a 200 km search region; improved star detection from 60% to 100%
  • Expanded triangle signatures from 20 to 364, reducing search space from 38M km² to 126k km²
Key Outcome: near a degree of localization accuracy from sky imagery alone, with a 99.7% reduction in search space. No GPS, no ground beacons, no surface landmarks.
Particle FilterAstropyOpenCVNumPyLocalization
View on GitHub →
Reinforcement Learning · Manipulation Nov 2025

Hierarchical RL for Robust Robotic Control under Uncertainty

A decoupled policy architecture for long-horizon manipulation that remains stable under significant perception noise.

  • Decoupled policy stack: Soft Actor-Critic (SAC) for high-level goal selection, PPO for low-level force control, outperforming monolithic methods by 5x on long-horizon tasks
  • Modeled as a POMDP to handle observation noise up to 40% object half-length, mitigating compounding errors
  • Custom dense reward function with stability heuristics and multi-stage curriculum for robust sim-to-real transfer
Key Outcome: 2x average stack height vs. discrete baselines (27.18 vs 14.85 blocks) in PyBullet.
SACPPOPyBulletPOMDPPython
Full-Stack Autonomy · ROS2 Sep 2025

Autonomous Ground Robot: Full-Stack Navigation Stack

A complete on-robot autonomy system for indoor navigation in cluttered, dynamic environments with moving obstacles.

  • Modular autonomy stack in ROS2/C++: sensor fusion via EKF and global localization via Particle Filter (MCL)
  • YOLO-based perception integrated with Occupancy Grid Mapping to inform a hybrid planner (A* global + DWA local)
  • Trajectory tracking controller minimizing cross-track error, validated through real-world and simulation testing
Key Outcome: Full autonomous navigation validated in dynamic environments, with robust sensor fusion, perception, and hybrid path planning integrated end-to-end.
ROS2C++GazeboEKFSLAMYOLO

03. Experience

Autonomy Intern

Archer Aviation Jun 2026 – Aug 2026 San Jose, CA
  • Built a multi-agent planner for probabilistic search that uses aerial imagery to plan swarm-level search paths, from idea to shipped in 10 weeks.
  • Plans in seconds with near-optimal search performance and swarm-level safety and deconfliction guarantees.
  • Supported ground and flight testing of the autonomy stack.

Student Researcher

Stanford Multi-Robot Systems Lab Jan 2026 – Present Stanford, CA
  • Leading AORA, an agentic orchestrator for closed-loop autonomous policy training, and designing visuomotor policies for language-guided drone navigation. See Research.

Engineering Intern, Software Tools

Boeing May 2024 – Jul 2024 Chennai, India
  • Built Python automation tooling for 787 payloads supplier transition: 30% faster design release.

Vehicle Systems Engineering Intern

Tata Motors Ltd. Sep 2023 – Dec 2023 Pune, India
  • Designed and validated aluminium chassis for electric ILCV trucks, up to 35% weight reduction; selected for prototyping and advanced to New Product Development.

Team Captain

Team Veloce, University UAV & Robotics Team May 2023 – May 2024 Vellore, India
  • Directed a 25-member engineering team across the full hardware-software lifecycle, fielding 7 UAVs and ground robots for competition and field testing.

Aeromodelling Engineer

Team Veloce, University UAV & Robotics Team Apr 2022 – Apr 2023 Vellore, India
  • Designed, built and flight-tested fixed-wing and multirotor airframes for national competition entries.

04. Skills

Deep Learning

VLAsVisuomotor PoliciesFoundation Models VLMsKnowledge Distillation Imitation Learning (BC, DAgger) Reinforcement Learning (TD3+BC, SAC, PPO) Transformers / Cross-AttentionPyTorch

Computer Vision

Sim2Real3D Gaussian Splatting Monocular Depth EstimationVisual Localization SLAMSensor Fusion (LiDAR/RGB-D/Vision) Detection & Multi-Object Tracking (YOLO, DeepSORT)

Robotics & Autonomy

Agentic OrchestratorsMulti-Agent Swarm Planning Optimal Control / MPC State Estimation (EKF/IMM, Particle Filter) GPS-Denied NavigationPath Planning (A*, θ*, DWA)

Tools & Infrastructure

PythonC++ROS2NumPy DockerGitLinux MuJoCoCARLAGazebo

05. Background

Education

Stanford University

MS in Aeronautics & Astronautics  ·  GPA 4.0 / 4.0

Sep 2025 – Mar 2027

Robot Autonomy (Multi-Robot Systems) · State Estimation · Deep Reinforcement Learning · Computer Vision

Vellore Institute of Technology

BS in Mechanical Engineering  ·  GPA 4.0 / 4.0

Jul 2021 – Jul 2025

Graduated top of class · Merit Scholarship '22–'25 · Student Achiever '24 & '25

Fellowships

Fellowship

Narotam Sekhsaria Foundation Fellow

One of India's most selective postgraduate fellowships, for graduate study abroad

2025
Fellowship

KC Mahindra Scholar

National fellowship, KC Mahindra Education Trust

2025
Scholar

JITO JEAP Scholar

Jain International Trade Organization postgraduate fellowship

2025

Competition Wins

1st

GKN Aerospace Sustainable Aviation Challenge

National winner. INR 300,000 prize, autonomous eVTOL systems integration

Nov 2023
1st

Forbes Marshall Technical Challenge

Ranked 1st of 200+ participants nationwide

Mar 2024
2nd

Honeywell Aerospace Ideathon

National runner-up. INR 50,000 prize

Mar 2023
3rd

AIRBUS National Flight Challenge

Top 3 nationally. Dynamic braking safety system for inclined runways

Feb 2024

Publications & Patents

Earlier work in aerospace materials and mechanical systems, from my undergraduate research.

Journal · Grain Boundary Evolution in GRCop Alloys. Results in Engineering, Elsevier (2025) Journal · Aerodynamic Tail Configuration Study (2025) Patent · Wear Analysis Equipment for Marine Applications, No. 202541056455 Patent · Electronic Spherical Steam Trap, No. 427140-001

Get In Touch

If you're building in autonomy or robot learning, let's talk.

kothari1@stanford.edu LinkedIn GitHub