I'm a master's student in Computer Science at Tufts University, currently researching safe and performant autonomous robotics in the [SPARC Lab](https://www.ryancosner.com/) with Prof. Ryan Cosner. My work focuses on learned Control Barrier Functions (CBFs) for legged robots — how to make policies that stay safe under disturbances and distribution shift without giving up task performance.

Previously I worked in the [MULIP Lab](https://www.eecs.tufts.edu/~jsinapov/) with Prof. Jivko Sinapov on multi-modal tactile perception and robot skill learning.

**Research interests:** multimodal learning, autonomous systems, robot skill learning, world models (especially JEPA-style approaches).

**Before Tufts:** BS in Computer Science from Brandeis (Magna Cum Laude, May 2024), where I also played NCAA Varsity Baseball. Industry experience at Wealth.com (Software Engineer, backend + security tooling) and HealthFlexx (LLM-based assistant for medical devices, prototyping RAG pipelines).

**In my free time:** basketball, reading philosophy, and practicing MMA.

---
layout: archive
title: "CV"
permalink: /cv/
author_profile: true
---



My full CV is available as a PDF: [Download CV (PDF)](/files/Chris_Liang_Resume.pdf).

## Education
- **MS in Computer Science**, Tufts University (expected May 2027)
- **BS in Computer Science**, Brandeis University (May 2024, *Magna Cum Laude*)

## Research Experience
- **SPARC Lab, Tufts University** — Graduate Research Assistant (Jan 2026 – Present). Advisor: Prof. Ryan Cosner.
- **MULIP Lab, Tufts University** — Graduate Research Assistant (Aug 2025 – Jan 2026). Advisor: Prof. Jivko Sinapov.

## Professional Experience
- **Wealth.com** — Software Engineer (Jan 2025 – Jul 2025)
- **HealthFlexx** — Applied Scientist Intern (Jun 2023 – Aug 2023)

---
layout: archive
title: "Research"
permalink: /research/
author_profile: true
---

## SPARC Lab — Safe + Performant Autonomous Robotics and Control
*Tufts University · Jan 2026 – Present · Advisor: Prof. Ryan Cosner*

At the [SPARC Lab](https://www.ryancosner.com/) we work toward making robots you can trust — robotics research at the intersection of machine learning and control theory, aiming for provably safe and dynamic robot autonomy in real-world settings.

My current work focuses on learned Control Barrier Function (CBF) approaches that navigate the safety-performance tradeoff, adaptively modulating nominal controllers to remain robust under disturbances and distribution shift. I've been architecting and training policy distillation pipelines that produce adaptive safety parameters — bridging data-driven adaptability with formal control-theoretic guarantees, with an eye toward sim-to-real validation on legged robot platforms.

## MULIP Lab — Multi-Modal Learning, Interaction, and Perception
*Tufts University · Aug 2025 – Jan 2026 · Advisor: Prof. Jivko Sinapov*

I worked on multi-modal tactile data collection and transfer learning for robot skill learning. Specifically: building and executing tactile data collection protocols on a UR5 manipulator across a range of tool-use actions, developing pipeline components for data preprocessing and representation analysis, and maintaining broader lab infrastructure.

---
title: "Twist-RL: Reinforcement Learning for Helical Manipulation"
excerpt: "Extending force-based manipulation learning to helical articulation — training a robot-agnostic policy to unscrew bottle-cap-like objects by learning in object force-space."
collection: portfolio
---

**Course:** Reinforcement Learning, Fall 2025

We extended the FLEX (Force-based Learning for Extended Manipulation) framework to a new class of articulated motion — helical/screw joints — training a robot-agnostic policy that transfers across platforms by learning in object force-space rather than joint configuration space.

**What I built:**
- Designed the MDP: screw-axis state and axial torque action space, multi-term reward combining success, progress, and minimum-torque objectives.
- Constructed a custom helical joint in MuJoCo by coupling hinge and slide joints with an equality constraint.
- Trained a TD3 policy (actor-critic with clipped double Q-learning and delayed actor updates) for continuous control.
- Validated on a Franka arm in Robosuite, producing stable unscrewing behavior on a bottle model.

**Stack:** PyTorch, MuJoCo, Robosuite, Stable Baselines3.

---
title: "Probabilistic Trajectory Reconstruction from Keyframes"
excerpt: "Reconstructing high-frequency robot trajectories from sparse keyframes using Gaussian Processes and a masked variational autoencoder."
collection: portfolio
---

**Course:** Bayesian Deep Learning, Spring 2025

Robot trajectories recorded at high frequency produce large, bandwidth-heavy datasets. This project explored whether sparse keyframes could be used to reconstruct the full trajectory with calibrated uncertainty — useful for low-bandwidth control pipelines and uncertainty-aware motion planning.

**What we built:**
- **Baseline:** per-joint Gaussian Process regression over RBF and Matérn kernels, treating each joint independently.
- **Upgrade:** a masked variational autoencoder that encodes the entire trajectory with learned per-timestep mean and variance, using the keyframe mask to distinguish observed from imputed samples.
- **Evaluation:** compared both methods on MSE and credible-interval coverage across keyframe rates on a real SO-101 manipulator dataset (166 teleoperated block-picking episodes).

**Finding:** the VAE wins at sparse keyframe rates (≤ 1 Hz) by exploiting dataset-wide motion structure; the GP stays competitive at dense rates where local interpolation is sufficient.

**Stack:** PyTorch, NumPy, SciPy.

---
title: "Deep Document Unwarping"
excerpt: "Rectifying non-rigidly deformed document images with a Vision Transformer-based pipeline that predicts a dense UV coordinate map."
collection: portfolio
---

**Course:** Computer Vision, Fall 2025

Unlike planar document rectification which can be solved with a single homography, crumpled and folded pages require pixel-wise displacement estimation. This project designed a deep learning pipeline to rectify such non-rigid deformations.

**What was built:**
- Vision Transformer backbone to reason about 3D document geometry from a single crumpled input image.
- Prediction head outputting a dense UV map — a per-pixel lookup table into the flat, rectified document.
- Differentiable warping layer resampling the input image through the predicted UV map to produce the final output.
- Trained on a synthetic dataset of deformed documents (~8.6 GB).

**Stack:** PyTorch, OpenCV, Hugging Face Transformers.

---
title: "Brain Hemorrhage Detection with DDPM Reconstruction"
excerpt: "Using a diffusion model trained only on healthy brain scans to detect hemorrhages via reconstruction error."
collection: portfolio
---

**Course:** Generative AI, Fall 2025

We investigated whether a Denoising Diffusion Probabilistic Model (DDPM) trained solely on "healthy" images could be used for anomaly detection in medical imaging — the idea being that reconstructing an unhealthy image would produce a noticeably worse reconstruction than a healthy one.

**What we built:**
- Trained a UNet-based DDPM (64 base channels, channel multipliers (1,2,4), 1000 timesteps, linear β schedule) on clean images.
- Validated the approach on a synthetic MNIST corruption benchmark (AUROC **0.93**).
- Scaled to brain CT scans for hemorrhage detection, comparing a from-scratch ~30M parameter model against a pretrained ~114M parameter variant.

**Findings:**
- Diffusion models are decent at image-level anomaly detection but weaker at pixel-level localization.
- Pretrained models achieved lower validation reconstruction error but were actually worse at anomaly detection — a reminder that the training objective and the deployment task aren't always aligned.

**Stack:** PyTorch, Hugging Face Diffusers.

*Collaborators: Dennis Loevlie, Roger Brockett.*

---
title: "3D-Printed Legged Robot"
excerpt: "Designed, printed, and walked a custom legged robot driven by a Raspberry Pi control stack."
collection: portfolio
---

**Course:** Robotics, Fall 2025

Built a legged robot end-to-end — from CAD design of the mechanical parts through 3D printing, assembly, and the full control stack needed to make it walk.

**What we built:**
- Custom CAD parts for legs, body, and joint brackets.
- Assembly: motors, motor drivers, Raspberry Pi as the control MCU, power conversion electronics.
- Walking gait implementation on the physical platform.

**Stack:** Fusion 360 / CAD tooling, Python on Raspberry Pi.

*Collaborator: (partner led mechanical design).*