Badrikanath Praharaj CV

Badrikanath Praharaj

9 accepted papers, including ICRA 2026, RSS 2026 and NeurIPS 2025. Founding engineer at Kosha Labs, building datasets for trillion-dollar frontier AI labs. Robot learning intern at IISc Bangalore, turning real-world data into Isaac Sim digital twins for training robots.

B.Tech grad · open to PhD and research roles

Badrikanath Praharaj smiling, in a brown shirt, against wooden slats
01

Research

Robots are trained in made-up worlds and then fail in real ones. I build digital twins from real scans, and test robots on real hardware, so the gap shows up before deployment. Every paper has a short explainer video: press play on its figure.

Next: full papers. These are workshop papers. Over the coming year I am extending each one into a full paper for a main venue: ICRA, RSS, IEEE RA-L and T-RO.

  1. A real copper mine, rebuilt for robots to learn in.

    Full title and authors

    MineXplore: An Open-Source Reinforcement Learning Exploration Benchmark for GNSS-Denied Underground Environment

    Abhishek S*, Badrikanath Praharaj*, Sreeram M.V.* * equal contribution

    Mining RL had no benchmark. We built an open-source MuJoCo simulation of a real Chilean copper mine with a contour-to-MJCF pipeline.

    • 95.38%geometric IoU
    • 88.9%PPO coverage
    • ICRA 2026 · Xplore workshop
    • arXiv
    • PDF
  2. 351 metres of real metro tunnel, and no GPS.

    Full title and authors

    Subterra: An Open-Source Validated Benchmark for GNSS-Denied Tunnel Navigation

    Badrikanath Praharaj, Abhishek S, Sreeram M.V.

    Tunnel navigation had no RL testbed with real geometry, so its difficulty was invented. We rebuilt one from LiDAR scans and checked it against 199 metrics.

    • 97.1%fidelity, 199-metric suite
    • 2.46 cmsurface error
    • 90.02%PPO coverage
  3. A cheap robot that notices when its LiDAR is wrong.

    Full title and authors

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

    Abhishek S*, Badrikanath Praharaj*, Sreeram M.V.* * equal contribution

    Cheap LiDARs are fooled by glass and mirrors without noticing, and SLAM quietly corrupts the map. SENTINEL checks each scan against depth from a second sensor.

    • <$300robots
    • 5failure modes
    • 0training needed
  4. How long will the robots really take?

    Full title and authors

    CADENCE: Predicting Realized MAPF Execution Time Beyond Sum of Costs

    Abhishek S*, Badrikanath Praharaj*, Sreeram M.V.*, Mohan Prakash V * equal contribution

    Multi-robot plans are scored by sum of costs, which is informative but incomplete about how long they take on real robots. Adding a plan's motion burden cuts the prediction error sharply.

    • 480hardware trials
    • 7robots
    • 48.6–59.8%lower held-out MAE
  5. Ask the model why a photo looks fake.

    Full title and authors

    REVEAL: Reasoning and Evaluation of Visual Evidence through Aligned Language

    Ipsita Praharaj, Yukta Butala, Badrikanath Praharaj, Yash Butala

    Fine-tuned forgery detectors fail to generalize. REVEAL reframes detection as zero-shot VLM prompting, with eight-factor scoring and 3×3 region grids.

    • 0.39→0.85Gemini F1, Columbia (+118%)
    • 0.92GPT-4.1, rivaling baselines
    • NeurIPS 2025 · WiML workshop
    • arXiv
    • PDF

Also published

02

Projects

Seven builds, from swarms in simulation to servos on a bench. Each card opens a short explainer with real footage where it exists; drawn scenes are labelled as illustrations.

Capstone · VIT Bhopal · with Sreeram M.V.

MineSwarm

Three LiDAR robots learn to cross a GPS-denied mine with no central controller. They reach 91% goal success across an 8-phase curriculum, and transfer zero-shot to real hardware.

MADDPG + TD3 · PyBullet · ROS 2 / Gazebo

Code

IISc Bangalore

PowerSplit

Separate batteries for thinking and for driving, each sized to its own load. A three-robot team ran 74 minutes instead of 65 (+13.5%).

Raspberry Pi · coulomb counting · velocity control

Design files

IISc Bangalore

ROSMASTER

A Yahboom X3 set up as a ROS 2 perception testbed, with a RealSense D435 depth camera and a Livox Mid-360 LiDAR for a full 360° view.

ROS 2 · RealSense D435 · Livox Mid-360

Specs & media

VIT Bhopal · 2024

HexaGrip

A 3-DoF servo arm and a companion robo-car, built and wired by hand, and shown live at the Industry Conclave 2024 Project Expo.

Arduino · PCA9685 · HC-SR04

Code

VIT Bhopal

Road-object detection

YOLOv8 trained on BDD100K and Roboflow data to find cars, two-wheelers, buses and pedestrians in mixed traffic.

YOLOv8 · BDD100K · Roboflow

Code

VIT Bhopal

Clima-AI

An Arduino weather station built around ten sensors. A captured log in the repo shows six channels, from light and CO to rain and humidity.

Arduino · 10 sensors · SD logging

Repo

VIT Bhopal

WeedSense

A weed-detection robot design: Mask R-CNN segments weeds, LiDAR keeps it in the row (simulated in ROS 2 Gazebo), and readings go to ThingSpeak.

Mask R-CNN · LiDAR · ROS 2 Gazebo

Code not public yet

github.com/Badri1401 More builds, notebooks and course projects Open GitHub →
03

Work

  1. Kosha Labs, founding engineer

    Sep 2026 – now

    Video-data quality for datasets used by OpenAI and Google DeepMind. Raised QC-passed video by 23.5%, and built a gyro-plus-image blur detector that doubled recall (34% to 68%) with 59% fewer false rejects.

  2. IISc Bangalore, HiRo Lab, robot learning intern

    Jun 2026 – now

    Extending a real-to-sim pipeline that turns real-world datasets into Isaac Sim digital twins for RL (2 environments built), and training an exploration policy for uncertainty-aware SLAM with motion-capture DAgger. With Prof. Ravi Prakash.

  3. IISc Bangalore, Multi-Agent Systems (MAS) Lab, control and optimization intern

    Aug 2025 – Jun 2026

    A real-time, conflict-free coordinator for robot teams, scalable to 10 robots, tested in 125+ OptiTrack experiments with 0.25 s schedule error and 0.12 s inter-robot wait. With Prof. Pavankumar Tallapragada.

  4. Intel, LLM inference and optimization intern

    May – Sep 2024

    Cut T5-small CPU latency by 30% (120 to 84 ms) with OpenVINO quantization and operator fusion, then fine-tuned it with LoRA for summarization, raising ROUGE-L by 4.2%.

04

About

I've been building robots since my second year of college, with one goal: automate the physical world, starting with its hardest places.

What I enjoy most is owning the whole stack: the simulator, the controller, the policy, and the hardware that has to survive contact with reality. My bias is simple. I'd rather ship a benchmark someone can break than a number nobody can reproduce.

Away from the lab, I go looking for rough terrain myself: hiking, trekking up mountains, and riding Karnataka’s back roads with a camera in the bag. It’s usually the kind of ground I’d like robots to handle one day.

  • EducationB.Tech, Electronics and Communication (AI and Cybernetics), VIT Bhopal, 2022–2026. CGPA 8.9, a perfect 10.0 in the final year, and first in the department in three of four years.
  • ReviewingNeurIPS 2026 WiML workshop (robotics track) and MEXAAIDC.
  • CommunityCo-founded Eureka, a student R&D club, and guided 7+ papers and a patent.
  • HackathonCommunity Favorite (55+ public votes) at the LycheeAI × REVEL Isaac Sim hackathon.
  • ToolsMuJoCo, Isaac Sim, PyBullet, ROS 2, OptiTrack, LiDAR and RGB-D, PyTorch.

On the road

Badrikanath leaning on his red sport motorcycle on a tree-lined trail Red sport motorcycle parked on a forest trail The motorcycle on red earth beside a lake, hills behind

Through my lens

A ginger cat looking straight up at the camera A yellow building corner against a blue sky with power lines Trees and sky reflected in a flooded paddy field
Concrete pillars and a coffered ceiling reflected in a polished floor A brown puppy resting on a mossy ledge
05Contact

Working on robots for the real world? Let's talk.

PhD openings, research collaborations, or a robot that keeps failing in a way simulation didn't predict. The form opens your email app with the message filled in.

Or write directly: badrikanath.praharaj@gmail.com

Explainer

No sound needed: the captions are part of the video. Arrow keys skip 5 seconds.