A CoRL 2026 submission on VLM-guided reward compilation for VLA policies.
VLA Models / Reinforcement Learning / Tactile World-Action Models
Haodi (Woody) Hu
Ph.D. graduate from the University of Southern California currently focused on vision-language-action models, tactile-grounded world-action models, reinforcement learning, and dexterous manipulation systems for robust embodied autonomy.
I am currently a Senior Machine Learning Engineer at Sandisk. Previously, as a Research Scientist Intern at MERL, I developed ReCoVLA for VLM-guided residual RL recovery and TacSushi for tactile-grounded world-action control with a dexterous hand. My broader research grew out of USC RoboLAND and spans robot learning, embodied reasoning, granular loco-manipulation, and multi-robot coordination, where I worked with Professor Feifei Qian and collaborated with Professor Daniel Seita.
Dexterous hand manipulation, tactile prediction, risk-constrained control, and policy adaptation.
Research Direction
Building VLA, tactile world-action models, and RL systems for robot manipulation
My current research connects language-conditioned robot policies, reinforcement learning, and tactile-grounded world-action modeling for manipulation in unstructured settings, spanning long-horizon mobile autonomy and contact-rich dexterous hands.
Core areas
Vision-Language-Action Models
Reinforcement Learning
Mobile Manipulation
Tactile World-Action Models
Long-Horizon Manipulation
Robot Learning
Embodied Reasoning
Policy Adaptation
Multi-Robot Systems
Recent Updates
News and momentum
Joined Sandisk as a Senior Machine Learning Engineer working on RL optimization and VLM post-training.
Submitted ReCoVLA, a VLM-guided reward compilation framework for VLA policies, to CoRL 2026.
Research Scientist Intern at MERL focusing on VLA recovery, reinforcement learning, and tactile-grounded world-action models.
Paper accepted to the 9th Conference on Robot Learning: Granular Loco-manipulation.
Worked as a Machine Learning Engineer intern in the Data Science group at SanDisk.
Co-organized the 5th workshop on representations and manipulating deformable objects at ICRA 2025.
Presented research on legged robot loco-manipulation and obstacle-aided locomotion at ICRA 2025.
T-RO paper accepted on obstacle-aided trajectory control through sequential gait composition.
Selected Publications
Recent projects, papers, and videos
Recent work across VLA models, reinforcement learning for manipulation, dexterous tactile manipulation, granular loco-manipulation, obstacle-aided navigation, and multi-robot systems.
Spotlight paper
ReCoVLA: VLM-Guided Reward Compilation for Failure Recovery in Vision-Language-Action Policies
A VLA/RL framework that keeps the base policy frozen, uses Qwen3-VL to infer structured task context, and compiles stage-gated rewards. ReCoVLA improved simulation success from 36.7% to 66.7% and achieved 61.7% success on a physical Fetch robot.
TacSushi: Learning Sushi Making with a Dexterous Tactile Manipulator and World Models
An emerging Shadow Hand project on tactile sensing, teleoperation, and dexterous manipulation of deformable objects for sushi-making tasks, including hand-tracked teleoperation, soft-object handling, and cake manipulation.
Granular Loco-manipulation: Repositioning Rocks Through Strategic Sand Avalanche
A research direction in which legged robots use learned models of granular dynamics to reshape terrain and indirectly manipulate obstacles.
Learning Granular Media Avalanche Behavior for Indirectly Manipulating Obstacles on a Granular Slope
A learning-based framework for predicting granular avalanche behavior to support indirect manipulation on sandy slopes.
Method for Detecting Micron Cracks on a Magnetic Rotor Surface Based on a Support Vector Machine
Teaching, Mentoring, and Service
Academic activities beyond publications
The new structure makes this material easier to scan than a long text block, while still preserving the substance from your current site.
Teaching experience
- Robot Mobility (EE599) — Teaching Assistant, Fall 2022
- A Computational Introduction to Deep Learning (EE541) — Teaching Assistant, Spring 2023, Fall 2023, Spring 2024
- MOS VLSI Circuit Design (EE477L) — Teaching Assistant, Fall 2024, Spring 2025
Awards and service
- IEEE Access Exceptional Reviewer Recognition, 2026
- USC Viterbi CURVE Mentor Award, 2022–2025
- USC Viterbi Ph.D. Student Fellowship Award, 2021
- NEFU Excellent Graduates Award, 2019
- Workshop co-organizer for ICRA 2025 deformable objects workshop
Mentoring
- Supervised students including Luke Cortez, Jerry Wu, Seojoon Kwon, Tian Xie, and Brendon Lee.
- Served as Ph.D. mentor in the USC CURVE program across multiple cohorts from 2022 to 2025.
- Supported student researchers through project mentorship and conference preparation.
