NeurIPSPending
Mechanistic Interpretability and Steering of VLA Models Through Sparse Autoencoders
I co-authored research on understanding and steering Vision-Language-Action models through sparse autoencoders.
Purdue CS / Machine Intelligence / robotics researcher
I research machine learning and perception for robots.
I am a Computer Science student in Machine Intelligence at Purdue. My work is mostly in RGB-D perception, SLAM, Vision-Language-Action models, whole-arm manipulation, WAMs and closed-loop control.
MePranav Kumar
Publication
Field Log
Sim-to-real not too real; I work on making it real.
AMD robotics demo
Humanoid robotics showcase.
Purdue Robotics Day
Demoing VLA and classical manipulation hybrid research.
Vibe Robotics
Built and trained in house (literally).2× speed
SPOT playing Connect 4
Programmed SPOT to play Connect 4 as part of a data collection pipeline for policy training.
Connect 4 with Pi0.5
VLA Policy model demo for game interaction and embodied decision-making.
Research & Work
What I am working on
Vibe Robotics
Real-time perception and hybrid control for humanoid robots
- Robotics Engineer Intern — RGB-D perception, point clouds, and SLAM-based spatial modeling for humanoid interaction.
- Build real-time perception pipelines that turn raw sensor streams into structured scene representations.
- Develop hybrid control systems combining teleoperation with autonomous policies for closed-loop execution.
SCALE Robotics Lab
Vision-Language-Action models for closed-loop robot control
- Trained VLA models (Open-Pi, OpenVLA, ACT) for closed-loop robotic control from multimodal inputs — imitation learning and RL for policy refinement.
- Built perception-to-action pipelines integrating RGB-D sensing, SLAM, and state feedback for real-time inference.
- Conducted sim-to-real experiments analyzing generalization gaps, failure modes, and robustness of learned policies.
- Built a real-world Connect 4 robotic system combining VLA policies with classical control (Purdue Robotics Day demo).
- Co-authored research on mechanistic interpretability and steering of VLA models through sparse autoencoders (SAEs).
- Working on hallucination mitigation in video models for World Action Models.
Previous Experience
Previous experience
Dow
Agentic ML workflows for scientific analysis
- Build agentic ML workflows with LangGraph and DSPy for automated scientific analysis, instead of one-off prompts.
- Focus on reproducibility — traceable, repeatable model outputs that scale beyond a single analysis run.
Luna Social
Event pipelines and graph-based interaction modeling
- Software Engineer Intern — re-architected high-throughput event ingestion pipelines for product behavior data.
- Designed graph-based interaction models for large-scale prediction systems, bridging product engineering and ML.
Purdue Aerial Robotics Team
UAV perception and control systems
- Worked on sensor fusion pipelines for stable real-time UAV flight.
- Early robotics systems work before moving deeper into ML for robot control.
Projects
Selected robotics projects
robot perception
RealSense to Unity 3D Perception & SLAM Pipeline
I built a real-time RGB-D pipeline from an Intel RealSense D435i for 3D reconstruction, SLAM-based localization, and spatial mapping in Unity.
VLA + classical control
Connect 4 Robot Demo
I built a real-world Connect 4 robotic system that combined learned VLA policies with classical control for Purdue Robotics Day.
Stack
Tools I use
Languages
PythonC/C++JavaRJavaScriptTypeScriptSQL
Robotics / AI
ROSSLAMRGB-D PerceptionPoint CloudsSensor FusionVLAWAMsIsaacMuJoCo
ML
PyTorchTensorFlowReinforcement LearningImitation LearningComputer VisionLangGraphDSPy
Systems
DockerLinuxFlaskNode.jsReactMongoDBUnity
Signals
Recognition
HackMIT 2025 - Citadel Challenge Winner
NASA SUITS 2025 - Best Innovative
Catapult Hackathon 2025 - Runner-up
HackMIT 2024 - Top 12 / 250+
Boilermake XII - Top 8 / 109
CS Base Climate Hackathon 2024 - Advance Tier Second Place
Purdue Hello World Hackathon 2023 - 2nd place
Indian National Mathematical Olympiad - National Team