~/slam ~/3d-vision ~/tracking ~/vla

Jash Shah.

Robotics and ML researcher. I build systems that figure out where they are, what they see, and what to do about it: SLAM, multi-object tracking, and vision-language-action models.

github linkedin email
what you are watching: visual odometry drifts, a loop closure fires, iSAM2 snaps the whole trajectory back. on real KITTI seq 00 my pipeline takes ATE from 29.14 m to 7.76 m this way. click the map to close the loop yourself.
36
merged PRs across 12 repos
2
tools live in official HF LeRobot products
1
guest post on the GTSAM blog
3.92
GPA, MS in Robotics @ CMU
x1

work

2023 - now

Robotics Researcher

Nissan Advanced Technology Center, Silicon Valley
  • Graph-based transformer pipeline for multi-object tracking: 0.71 AMOTA with real-time inference on vehicle
  • Sparse4Dv3 LiDAR-camera fusion tracker on nuScenes: 0.650 AMOTA, +0.17 over camera-only baseline
  • First end-to-end LiDAR+camera tracker (raw sensors to tracks, no separate detector), competitive with two-stage SOTA
  • CUDA-accelerated RANSAC ground-plane detection at 10 Hz; lag-compensated EKF fusion improving pose estimates ~43%
2023 - 2024

ML Satellite Researcher

Robotics Exploration Lab, Carnegie Mellon University
  • Led a team of 5 building two-stage region classification and landmark detection for low-power satellite edge hardware
  • 18 lightweight YOLOv8n landmark detectors: localization error under 50 km from 700 km altitude
2023 - 2024

Computer Vision Research Assistant

Norilla — Mixed Reality Learning Platform, Pittsburgh
  • OpenCV sunlight-compensation and temporal filtering for a vision system reliable at 30 FPS under changing light
  • Occlusion-aware background segmentation from Intel RealSense depth; k-NN classifier +27% in occluded scenes
2021

Computer Vision & Deep Learning Intern

RapidAI Vision, Defence Research & Development Organisation (DRDO)
  • Vision-based precision landing of a UAV swarm from 9 m using fiducial marker pads; 4% pad-width error in PX4/Gazebo SITL
  • Deployed a custom ORB-SLAM2 variant reconstructing the site to 90% accuracy from a monocular camera in real time
2020

Computer Vision Intern

Autonomous Robots & Multi-Robot Systems Lab, IIT Bombay
  • SqueezeSeg-style CNN + recurrent CRF for real-time road-object segmentation on LiDAR point clouds, IoU 0.65
  • Built a spherical-projection data loader mapping 3D point clouds to 2D for the network
2019 - 2020

Founder & Team Lead — Hyperloop

loopMIT, official Hyperloop team of MIT Manipal
  • Founded and led a cross-year team (freshers to PhDs); secured university funding to design a 6-ft Hyperloop pod
  • The team I founded went on to win the 2025 Global Hyperloop Competition (IIT Madras): 1st in Pod Demonstration + Best Mechanical Subsystem Design
2017 - 2019

Robotics Engineer

RoboManipal, official robotics team of MIT Manipal
  • Holonomic robot with a 3-DOF arm to solve Tetris — All India Rank 2; a kinematically-analysed quadruped leg
  • Maze-solving / line-following and omni-wheel bots for national competitions (IIT Kharagpur, ABU Robocon)
2016 - 2019

Automation & Mechanical Intern

Mahindra & Mahindra · Technofour Electronics · Saisons
  • End-to-end automation of connecting-rod deburring at Mahindra; SolidWorks conveyor + CNC code optimisation at Technofour
x2

projects

kf 01 | featured on the GTSAM blog

gtsam-splatfactors

3D Gaussian Splatting SLAM with a factor graph backend: gsplat rendering expressed as GTSAM factors, giving loop closure, multi-sensor fusion, and pose uncertainty that pure gradient-descent 3DGS pipelines lack. Up to 85% ATE reduction on KITTI with loop closure. Written up in 3D Gaussian Splatting meets Factor Graph SLAM on the GTSAM blog.

live in official HF tooling

lerobot-doctor

Dataset-quality diagnostics for LeRobot. Ships as the Doctor tab of the official LeRobot Dataset Visualizer, and on PyPI.

listed on awesome-3DGS

vggt-factor-refinement

Factor-graph refinement of VGGT for long-video pose estimation. 70% ATE improvement over naive stitching, COLMAP-free.

YC hackathon winner

zero-shot semantic mapping

VLM embeddings fused with a SLAM pose graph for language-conditioned mobile manipulation. Grand Prize + Best Use of Data at RoboHacks @ YC.

debugged a blind VLA

vla-lite

Runs OpenVLA-7B on an 8 GB GPU and fixes a silent bug that left the policy blind on modern transformers — same action for every image.

ranked 18th · Alphabet challenge

robotic cable insertion

Vision-guided cable insertion (sub-5 mm) for Intrinsic's AI for Industry Challenge; cut target-pose error 7×. Team HexDex.

3d reconstruction

recon3d

One-command 3D reconstruction from a single handheld video. No COLMAP or manual calibration.

also built
gtsam-torch — PyTorch autograd bridge for GTSAM factor graphs.
lerobot-bench — statistically rigorous A/B benchmarking for robot policies.
PointSLAM++ — SuperPoint + keyframe re-anchoring for dense RGBD SLAM.
CubeSat vision — lightweight CNN for spacecraft pose from ~700 km (CMU).
vision-based obstacle avoidance — TD3 RL fusing an RGB camera with 2D LiDAR.
hand tracking & handover — MediaPipe hand pose driving a Franka arm.
x3

open source

Reviewer at HuggingFace LeRobot. Every edge below is a repo I have merged code into; node size is merged PR count. Click a node for the receipts, or see the full list on GitHub.

GTSAM 11 LeRobot 9 PyTorch RL 3 Rerun 3 GTSAM Blog 2 RSL_RL 2 JAX 1 Transformers 1 Accelerate 1 wrap 1 LeRobot Visualizer 1 awesome-3DGS 1
x4

awards & recognition

2026

Grand Prize + Best Use of Data

RoboHacks @ Y Combinator, out of 120 builders.

2026

Featured on the GTSAM blog

Guest post: "3D Gaussian Splatting meets Factor Graph SLAM" on Frank Dellaert's GTSAM.

2026

Ranked 18th — Intrinsic AI for Industry Challenge

Alphabet's Intrinsic cable-insertion challenge. Team HexDex.

2018

All India Rank 2 — World Robot Olympiad

Also Rank 2 at the Zonal Autonomous Robotics Challenge.

2016

All India Physics Topper

Cambridge AS-Level Physics, India topper for the session.

2021

Drone Dojo scholarship

"Advanced Precision Landing" scholarship (~$400) for helping others through technical issues.

2024

Best Entrepreneur — Carnegie Mellon

Awarded during the MS in Robotic Systems Development at CMU.

2025

Global Hyperloop champions

loopMIT, the team I founded, took 1st in Pod Demonstration and Best Mechanical Subsystem at IIT Madras.

x5

education

2022 - 2024

Carnegie Mellon University

MS in Robotic Systems Development, GPA 3.92/4.0 · Pittsburgh, PA
2017 - 2021

Manipal Institute of Technology

B.Tech Mechatronics, Minor in Robotics & Automation, GPA 8.78/10 · Manipal, India
2015 - 2017

Aditya Birla World Academy

Cambridge A-Levels (Physics, Chemistry, Maths: A, A, A) · India Physics topper
2012 - 2015

Hill Spring International School

IGCSE · 5 A*s, 3 As
x6

beyond the lab

robotics3d visionhyperlooptrekkingscubacricketlego
loop closure detected: x7 -> x0 | trajectory corrected