Skip to content

Tianshu Huang / Research

My past and current work spans a wide range of topics in machine learning, including foundation models (ICCV '25), statistical learning (MLSys '25), NeRF-style neural-implicit inverse rendering (CVPR '24), and meta-learning (ICLR '22). I also actively support systems researchers by providing machine learning, statistics, and data science expertise (SIGGRAPH '26, OOPSLA '25, RTAS '25, EuroSys '25), while also supporting machine learning researchers working with real-world sensors and systems (3DV '26).

Currently, I'm focused on scaling and building out foundational models for low-level spectrum-based radar perception, as well as establishing an ecosystem and community for radar research more broadly via the RadarML open source initiative and Radar Interest Group @ CMU.

Publications

Implicit Surface Compression — with Good Old Discrete Cosine Transform and Motion Compensation
Tao Jin, Shengxi Wu, Tianshu Huang, Mallesham Dasari, Srinivasan Seshan, Anthony Rowe
ACM Transactions on Graphics•(SIGGRAPH 2026)

Popular representations for 3D geometry such as meshes and point clouds are spatially and temporally discontinuous, and thus difficult to compress. Using implicit surface representations which are spatially continuous and temporally stable enables a simple compression scheme based on classical video coding paradigms to achieve order-of-magnitude bitrate reduction.

Implicit Surface Compression — with Good Old Discrete Cosine Transform and Motion Compensation

RadarSim: Simulating Single-Chip Radar via Multimodal Neural Fields
Chuhan Chen, Tianshu Huang, Akarsh Prabhakara, Chaithanya Kumar Mummadi, Zhongxiao Cong, Anthony Rowe, Matthew O'Toole, Deva Ramanan
3DV 2026

While radar reconstructions can accurately capture radar-specific properties, they are fundamentally limited by the low resolution of the underlying sensor. We propose a unified differentiable renderer, RadarSim, which leverages the high angular resolution of RGB cameras to generate Doppler radar range images with accurate radar properties from a camera-initialized neural field.

RadarSim: Simulating Single-Chip Radar via Multimodal Neural Fields

Towards Foundational Models for Single-Chip Radar
Tianshu Huang, Akarsh Prabhakara, Chuhan Chen, Jay Karhade, Deva Ramanan, Matthew O'Toole, Anthony Rowe
ICCV 2025•Oral Presentation — Top 0.57% (64/11152) of all submissions

With enough training data, transformer models can extract 3D outputs of surprising quality from indexpensive, low-resolution single-chip radar. These models scale well with additional training data, and can generalize across different settings and be readily fine tuned for new tasks.

PaperRadarMLSlidesPosterICCV Talk

Towards Foundational Models for Single-Chip Radar

Unveiling Heisenbugs with Diversified Execution
Arjun Ramesh, Tianshu Huang, Jaspreet Singh Riar, Ben Titzer, Anthony Rowe
OOPSLA 2025•Artifact Available — Artifact Functional — Results Reproduced

Heisenbugs - bugs which change their behavior under observation - are among the toughest challenges when debugging programs. We propose to catch these bugs using stochastic instrumentation, and demonstrate the surprising effectiveness of debugging across a diverse set of platforms.

GithubInstrumentation Framework

Unveiling Heisenbugs with Diversified Execution

Interference-Aware Edge Runtime Prediction with Conformal Matrix Completion
Tianshu Huang, Arjun Ramesh, Emily Ruppel, Nuno Pereira, Anthony Rowe, Carlee Joe-Wong
MLSys 2025•Artifact Available — Artifact Functional

Performance prediction in heterogeneous systems such as edge computing is best formulated as matrix completion, which can be extended to handle complex, edge-specific concerns such as interference and uncertainty quantification.

GithubZenodoSlides

Interference-Aware Edge Runtime Prediction with Conformal Matrix Completion

Silverline: Lightweight Virtualization and Orchestration of Distributed Systems
Arjun Ramesh, Tianshu Huang, Emily Ruppel, Dakshina Dasari, Behnaz Pourmohseni, Fedor Smirnov, Marco Giani, Paolo Pazzaglia, Charles Shelton, Nuno Pereira, Arne Hamann, Dirk Ziegenbein, Anthony Rowe
RTAS 2025

We develop Silverline: a framework for programming heterogeneous dynamic distributed systems. Using WebAssembly as a lightweight virtualization framework, we design silverline to be fault-tolerant and real-time, and show it in use on industrial-grade systems.

Silverline: Lightweight Virtualization and Orchestration of Distributed Systems

Empowering WebAssembly with Thin Kernel Interfaces
Arjun Ramesh, Tianshu Huang, Ben Titzer, Anthony Rowe
EuroSys 2025•Artifact Available — Artifact Functional — Results Reproduced

Webassembly (Wasm) adoption for new domains is often hindered by the lack of standard system interfaces. We propose directly exposing OS userspace syscalls as a compromise.

GithubRust Target

Empowering WebAssembly with Thin Kernel Interfaces

DART: Implicit Doppler Tomography for Radar Novel View Synthesis
Tianshu Huang, John Miller, Akarsh Prabhakara, Tao Jin, Tarana Laroia, Zico Kolter, Anthony Rowe
CVPR 2024•Oral Presentation — Top 0.78% (90/11532) of all submissions

We propose DART — Doppler Aided Radar Tomography, a Neural Radiance Field-inspired approach to radar novel view synthesis using implicit neural inverse imaging.

PaperImplementationData CollectionDataset

DART: Implicit Doppler Tomography for Radar Novel View Synthesis

Optimizer Amalgamation
Tianshu Huang, Tianlong Chen, Sijia Liu, Shiyu Chang, Lisa Amini, Zhangyang Wang
ICLR 2022

While meta-training optimizers from scratch is difficult, much attention has been directed towards developing a diverse body of analytical optimizers. Can we use these optimizers to meta-train a stronger optimizer?

PosterGithub

Optimizer Amalgamation

Patents

Active Method for predicting resource usage for applications in a distributed system
November 2025 • US 12,463,919 • Robert Bosch GmbH, Carnegie Mellon University

Pending Generation and use of synthetic radar views of a scene
Filed, October 2024 • WO2025131366A1 (WO DE) • Robert Bosch GmbH

Pending Method for detecting a memory access error in a multi-threaded application
Filed, September 2024 • US20250117309 (US JP KR DE) • Robert Bosch GmbH, Carnegie Mellon University

Pending Method for Predicting the Performance of a Software Program
Filed, March 2024 • US20250077389 (US CN DE JP) • Robert Bosch GmbH, Carnegie Mellon University

Pending Method for carrying out a decision for upgrading and/or deploying software on multiple heterogenous devices
Filed, December 2023 • US20240211241A1 (US CN DE) • Robert Bosch GmbH, Carnegie Mellon University

Conference & Invited Talks

Oral Towards Foundational Models for Single-Chip Radar
ICCV 2025 Main Conference Oral Presentation

Invited Towards Foundational Models for mmWave Radar
Invited Talk @ Bosch Research Sunnyvale, September 2025

Oral Interference-aware Edge Runtime Prediction with Conformal Matrix Completion
MLSys 2025, Conference Presentation

Invited The Radar Spectrum 2.0
2025 Signal Processing Colloquium @ Bosch Research • With Bosch Collaborators

Invited Towards Foundational Models for mmWave Radar
Invited Talks @ Bosch Research (January / September 2025), Bosch Mobility (February 2025)

Invited Grey-Box Program Analysis: Runtime Prediction and Beyond
Reliable Distributed Systems Tech Colloquium @ Bosch Research, October 2024

Oral DART: Implicit Doppler Tomography for Radar Novel View Synthesis
CVPR 2024 Main Conference Oral Presentation

Workshop Leveraging Wasm instrumentation
WebAssembly Research Day 2023 • With Arjun Ramesh

Workshop Giving the Cloud an Edge with WebAssembly
WebAssembly Research Day 2022 • With Arjun Ramesh