ROBOTICS & AI RESEARCHER

Hi, I'm Maxwell Ma

Graduate researcher at Rice University. I build intelligent robotic systems at the intersection of physics-informed learning, visuo-tactile perception, and sim-to-real transfer.

PROGRAM   ECE · Rice University
LOCATION   Houston, TX
RESEARCH   Manipulation · VLA · Motion Planning
INDUSTRY   3 yrs · Baidu AI
EMAIL       qm18@rice.edu
01

Research Projects

IROS 2026 · UNDER REVIEW Submitted Feb 27, 2026 · Manuscript #737

Vision–Tactile Mass Priors with Online
Physics-Informed Grip Optimization
for Compliance-Robust Manipulation

A staged grasp-and-lift framework performing online physical correction after contact: an IL base policy handles geometry-feasible approach; a vision model predicts mass priors pre-contact; a tactile micro-lift estimates realized mass and mismatch Δm; and a compact PINN outputs grip force by explicitly penalizing friction-cone violation, force-bound violation, and compliance-related pressure concentration. Implemented on a torque-controlled Franka Emika Panda with RGB-D + capacitive tactile arrays on Jetson AGX Orin.

1
Micro-lift realized-mass estimation. Short exploratory lift with quasi-static windowing estimates realized mass and prior–realized mismatch Δm as an interpretable diagnostic that exposes when the vision prior is unreliable.
2
Online physics-informed grip optimizer (PINN). Compact MLP trained with explicit penalty terms: Coulomb no-slip constraint, force-bound enforcement, and compliance deformation safety — one controller for both hard and soft objects.
3
Modular state-machine decomposition. Pre-contact observe → approach/close → micro-lift → lift/hold with online grip optimization → slip recovery. Geometry module is independently replaceable.
86.3% Success 6% Slip Rate 239 N Peak Force +9.3 pts vs best baseline −45 N vs Tactile Heuristic <0.5 ms / PINN step
Franka PandaRealSense D455Capacitive TactilePINNImitation LearningResNet-18PointNetROS 2Jetson AGX OrinPyTorch
RA · RobotII Lab · Rice University Aug 2025 – Present

LeRobot × Open-PI Integration —
Instruction-Conditioned VLA
for Tabletop Manipulation

End-to-end research pipeline for Vision-Language-Action policies targeting tabletop manipulation. The project spans the full ML development cycle: real-robot data collection → cross-dataset training → Isaac Sim evaluation → sim-to-real feedback iteration, bridging the LeRobot framework with Open-PI for scalable instruction-conditioned robotic control.

01
Data Collection

ROS 2 teleoperation pipeline recording synchronized RGB-D frames, proprioceptive states, and language annotations. Custom episode filtering and quality validation before dataset ingestion.

02
Training — LeRobot × Open-PI Adaptation

Adapted Open-PI dataset format for instruction-conditioned policy training within LeRobot. Multi-task training in PyTorch and JAX; cross-domain fine-tuning to bridge internet-scale data with lab demonstrations.

03
Evaluation — NVIDIA Isaac Sim Benchmarking

Systematic benchmark of open-source VLA models across standardized tabletop tasks. Metrics: success rate, trajectory length, latency, and language-instruction following accuracy under varied scene configurations.

04
Feedback Iteration — Sim-to-Real Gap Analysis

CI-backed architecture with reproducible training/eval scripts. Systematic failure-mode analysis; targeted data augmentation and domain randomization to close the sim-to-real gap iteratively.

LeRobotOpen-PINVIDIA Isaac SimROS 2PyTorchJAXRT-1 / RT-2OpenVLADomain RandomizationCI Pipeline
06

Contact

Open to research collaborations, lab discussions, and opportunities in robotics & AI.

02

Work Demonstration

▶ LeRobot DEMO
🎬

Video file not found. Place it next to index.html:

lehome_fold_clothes.webm

LeRobot Demo — Fold Clothes Manipulation

Instruction-conditioned VLA policy (LeRobot + Open-PI) executing a cloth-folding manipulation task, evaluated in NVIDIA Isaac Sim and deployed over a ROS 2 action server.

LeRobot Open-PI Isaac Sim ROS 2 PyTorch
03

Industry Experience

Senior Strategy AI Algorithm Engineer

Sep 2022 – Aug 2025
Beijing, China
Baidu
Full training stack: domain-adaptive pre-training → SFT → RLHF (PPO) with comprehensive evaluation harness for legal-domain LLM
RAG pipeline — vector retrieval + semantic rerank + hybrid BM25: +60% answer quality, −80% p95 latency
Led cross-functional pod (infra, data, legal SMEs) from design docs through weekly A/B reviews and KPI tracking
Production-scale deployment via streaming indexing, caching, and quantization
04

Skills

Robotics & Planning
OMPLRRT*BIT*ROS 2Isaac SimFranka PandaTask-Motion PlanningPDDL
VLA & Deep Learning
LeRobotOpen-PIPyTorchJAXTransformersPPOSACStable-Baselines3
LLM / GenAI
Pre-trainingSFTRLHFDPORAGFAISSLangChain
Languages & Infra
C++PythonCUDABashDockerAWSFastAPITensorRT
05

Education & Honors

Master of Electronic & Computer Engineering
Rice University
Aug 2025 – Dec 2026
Bachelor of Electronic Information Engineering
Northwest Minzu University
Sep 2018 – Jun 2022
First-class Academic Excellence Scholarship (Rank 1/114)
🏆
1st Prize · 12th Lanqiao Cup National Competition
Top 0.5% · July 2021
🎓
First-class Academic Excellence Scholarship
Rank 1/114 · Fall 2020