CHANDRAN
Work

Systems, studies, and
interfaces for autonomy.

These are not meant to read as a generic project grid. They are receipts for a direction: real robots under physical constraints, embodied AI policies that need observability and steering handles, and interfaces that help humans understand complex systems.

2026

Alpine Valley: Berry-Harvesting VLA

Founder in Residence / Latent Robotics

Building an agri-robotics company led me to the deeper problem: controllable embodied AI.

At Alpine Valley, I worked on autonomous berry harvesting with SO-101 robot arms, SmolVLA, teleoperation datasets, HG-DAGGER, MuJoCo/RL experiments, foundation stereo, and end-to-end VLA training. The application was strawberry picking, but the pivot was technical: modern robot foundation models need better interpretability, steerability, and validation before they can become reliable deployment systems.

  • Fine-tuned SmolVLA on roughly 450 episodes for a physical test setup
  • Demonstrated over 90% picking success on a fake-plant setup with silicon stems and plastic strawberries
  • Explored teleoperation, policy training, HG-DAGGER, RL/MuJoCo, and classical decomposition around the learned policy
  • Came away focused less on agriculture itself and more on steerable, reliable embodied AI
LinkedIn overview ↗
Alpine Valley: Berry-Harvesting VLA detail
Alpine Valley: Berry-Harvesting VLA detail
Alpine Valley: Berry-Harvesting VLA detail
Alpine Valley: Berry-Harvesting VLA detail
Alpine Valley: Berry-Harvesting VLA detail
VLARobot LearningTeleoperationFounder in Residence
2026

VLA Interpretability for Strawberry Picking

Hugging Face technical blog

A supervision layer around a SmolVLA policy, built to make the rollout easier to inspect and steer.

After the berry-harvesting work, I wrote up the interpretability layer: unsupervised stage discovery, a left-to-right HMM, Viterbi smoothing, VLA hidden-feature probing, and a live rollout HUD. The point was not only to make the policy understandable, but to search for handles that could eventually make black-box robot behavior more steerable.

  • Used task stages as an interpretable structure over a learned manipulation policy
  • Compared runtime signals from proprioception and VLA hidden representations
  • Built the supervision layer as an observer first, not a controller
  • Framed metrics carefully as evidence for observability, not finished scientific proof
Read the Hugging Face blog ↗LinkedIn post ↗
VLA Interpretability for Strawberry Picking detail
VLA Interpretability for Strawberry Picking detail
VLA Interpretability for Strawberry Picking detail
VLA Interpretability for Strawberry Picking detail
VLA Interpretability for Strawberry Picking detail
VLA Interpretability for Strawberry Picking detail
InterpretabilitySteerabilitySmolVLARobot Learning
Hugging Face LeLab Contributions cover
2026

Hugging Face LeLab Contributions

Open-source robotics tooling

My first serious open-source robotics contributions.

I have started contributing to LeLab, Hugging Face's browser UI around LeRobot workflows. The work spans bug fixes, features, setup, and usability improvements around robotics experimentation. I see this as part of the same thesis: if embodied AI is going to spread, the tooling has to make robots easier to configure, inspect, and iterate on.

  • Around six PRs merged, with more ongoing
  • Contributed to bug fixes, features, robot setup, and workflow usability
  • Open-source contribution, not employment or official affiliation with Hugging Face
  • A practical entry point into the tooling layer around embodied AI
View LeLab ↗
Open SourceLeLabLeRobotRobotics Tooling
FitCheck cover
2026

FitCheck

Build Small Hackathon / Hugging Face Space

A hardware-understanding tool for deciding what AI workloads a machine can realistically run.

FitCheck came from the same instinct outside robotics: make complex systems easier to reason about. The project helps users connect a hardware setup to likely AI workloads, exposing memory, inference, quantization, and deployment constraints in a more approachable way.

  • Built a Hugging Face Space and technical writeup for the Build Small Hackathon
  • Focused on memory limits, inference constraints, quantization, and practical deployment tradeoffs
  • A non-robotics example of reducing cognitive friction around AI systems
Read the FitCheck blog ↗
HardwareAI DeploymentHugging Face SpaceTechnical Writing
Multi-LLM Interface for Supermarket Robots cover
2024-2025

Multi-LLM Interface for Supermarket Robots

TU Delft MSc Robotics / Frontiers

A hierarchical conversational interface for helping people interact with service robots.

My master's work studied human-robot interaction through a multi-LLM conversational system for supermarket robots. It combined intent classification, retrieval-augmented generation, fine-tuning, speech interaction, and robot navigation integration. Looking back, it was an early version of the same question: how do humans supervise and communicate with robotic systems they cannot fully inspect?

  • Designed hierarchical routing between language models for different user intents
  • Combined RAG, fine-tuning, query classification, and spoken interaction
  • Connected language output to robot navigation and service-robot behavior
  • Published in Frontiers in Robotics and AI in 2025
Read the Frontiers paper ↗
HRILLMRAGService Robots
2025★ 1st place

Vision-Guided Pick-and-Place

FORGIS Physical AI Hackathon / Zurich

A fast prototype connecting vision, manipulation, calibration, and operator oversight.

At FORGIS, we built a vision-guided manipulation demo with a Dobot Nova 5: detect a pipe, pick it up, and place it into a box. We demoed it to Massimo Banzi, co-founder of Arduino. It was a compact example of the work I enjoy: build the perception-action loop, make the state legible, and get the robot to do the thing.

  • Trained an instance-segmentation model for pipe detection
  • Calibrated global-to-robot frame transforms for grasping
  • Executed grasp and placement behavior on a Dobot Nova 5
  • Demoed the final system live and won 1st place
LinkedIn post ↗
ManipulationComputer VisionAR/XRRapid Prototyping
Hyrox AR Coach cover
2025

Hyrox AR Coach

Snap x 3EALITY Hackathon

An AR fitness prototype with a burpee counter, coaching flow, and final pitch.

At the Snap x 3EALITY hackathon, I helped build an AR fitness prototype for Hyrox-style workouts. My main contribution was the burpee counter; I also worked with the team on the settings and product flow, then pitched the prototype at the end. It is supporting evidence rather than the center of my thesis: fast, camera-facing prototypes around feedback, embodiment, and human interaction.

  • Built the burpee counter for the AR workout flow
  • Contributed to settings and product flow with the team
  • Pitched the final prototype
  • A fast rep in camera-based interaction and human-facing feedback loops
LinkedIn post ↗Devpost ↗
AR/XRComputer VisionHuman-AI InteractionRapid Prototyping
Telosia Elderly-Care Robot cover
2025

Telosia Elderly-Care Robot

Junction Hackathon / Helsinki

A social robot prototype for elder-care assistance, built as a fast HRI proof of concept.

At Junction, I worked on an elderly-care robot concept and made the public project site. It belongs here as another rapid HRI prototype: taking a human problem, shaping the robot interaction, making the story legible, and getting a presentable demo together under hackathon constraints.

  • Built the public Telosia project site
  • Worked on the care-robot concept, pitch, and HRI framing with the team
  • Used the hackathon as a compressed test of robotics storytelling and prototyping
  • Part of the broader run of physical AI and human-robot interface projects
LinkedIn post ↗Project site ↗
Telosia Elderly-Care Robot detail
HRISocial RobotsRapid PrototypingRobotics
Unspool cover
2024

Unspool

Solo product / iOS + Android

A voice journaling app I designed, built, shipped, and got through both app stores myself.

Unspool was not a robotics project, and it did not become a huge consumer product. But it was important because I built the full thing end-to-end: voice capture, AI reflection, mood analysis, authentication, privacy, encryption, cloud sync, App Store review, and Play Store review. It proved I could take an idea all the way to a shipped product.

  • Designed and built the app single-handedly
  • Shipped across iOS and Android with under 100 early users
  • Combined voice capture, AI reflection, mood analysis, auth, and encrypted storage
  • Another expression of the same interest: AI interfaces that reduce cognitive friction
unspool.xyz ↗LinkedIn launch ↗
VoiceHuman-AI InteractionMobileProduct
Early Mechatronics Projects & Writing cover
2018-2022

Early Mechatronics Projects & Writing

Hackster / independent projects

Early robotics and mechatronics projects that collectively crossed 100k+ views.

Before the current embodied-AI direction became clear, I spent a lot of time building and writing about robotics and mechatronics projects. Those early posts were scrappier, but they were also proof that I liked making technical work public and useful for other people.

  • Published early robotics and mechatronics projects on Hackster
  • Reached 100k+ cumulative views across the account
  • Built the habit of documenting technical work in public
View Hackster ↗
MechatronicsTechnical WritingEarly Work