Long-horizon autonomy
RSI systems and context management support agentic, long-horizon autonomous research.
Zixing “Elwood” Lei · PhD Student, Shanghai Jiao Tong University · Google Scholar ↗
Agents that improve, explore, and act.
Supervised by Prof. Siheng Chen, I focus on Recursive Self-Improvement, Reinforcement Learning, and Embodied AI. I believe that using powerful agents to reshape today’s scientific and engineering R&D pipelines is a critical path toward a superintelligent future.
Three directions. One path toward autonomous intelligence.
Learning through exploration
RL environments and algorithms let agents explore autonomously and acquire efficient, learnable signals.
Verifiable physical intelligence
Embodied AI grounds agents in the physical world, where intelligence can be tested and verified.
Papers
No papers in this research direction.
2026
NavGPT-3: Harnessing Context in a Hierarchical Navigation Runtime
A hierarchical navigation runtime connects language-model reasoning with a low-latency action policy. Separate reasoning, acting, and monitoring threads manage context and control, enabling interruption handling and human-level success on RxR-CE.
Read paper ↗2026
IronMan: Information-Constrained Video-Action Learning for Robot Manipulation
An information-constrained video-action framework uses a dynamics-aware bottleneck to suppress irrelevant visual detail while retaining action-relevant cues, improving manipulation performance and robustness under out-of-distribution shifts.
Read paper ↗ Project page ↗2026
RSI-Master: Structuring Experiments to Guide Autonomous Model Improvement
An Experiment OS regularizes experimental actions and preserves traceable records, while reviewer-guided orchestration explores a growing research graph. Together they reduce hacking and strategy lock-in during autonomous model improvement.
Read paper ↗2026
Scaffolding Foundation Models into Physical-World Agents Pushes the Frontier of Long-Horizon Navigation
NavMCP scaffolds a VLM reasoning agent with a navigation foundation model through intent, observation, and memory, enabling persistent long-horizon embodied exploration without retraining and achieving state-of-the-art EQA and strong real-robot results.
Read paper ↗ Project page ↗2026
Ego2Robot: Scalable Robot Data Synthesis from Egocentric Human Data
A scalable pipeline that turns egocentric human manipulation videos into 18,561 hours of robot training data across 15 morphologies, improving out-of-distribution generalization for vision-language-action models.
Read paper ↗ Watch video ↗2026
Towards Long-horizon Embodied Agents with Tool-Aligned Vision-Language-Action Models
VLAs-as-Tools separates temporal reasoning from local physical execution, enabling high-level agents to plan, select specialized VLA tools, monitor progress, and recover over long horizons.
Read paper ↗2026
DataMaster: Data-Centric Autonomous AI Research
An autonomous data-engineering agent that searches, constructs, evaluates, and remembers data interventions to improve a fixed learning algorithm through downstream feedback.
Read paper ↗2026
PRL-Bench: A Comprehensive Benchmark Evaluating LLMs’ Capabilities in Frontier Physics Research
A benchmark that reconstructs end-to-end physics research workflows to evaluate long-horizon reasoning, autonomous exploration, and objectively verifiable scientific outcomes.
Read paper ↗2026
From Digital to Physical: Digital Agents as Autonomous Coaches for Physical Intelligence
EmboCoach-Bench evaluates agents that iteratively design, debug, and optimize embodied policies through environment feedback, spanning reward design and policy architecture.
Read paper ↗2025
EmboMatrix: A Scalable Training-Ground for Embodied Decision-Making
A scalable training ground combining large-scale task generation, distributed simulation, and multi-level rewards so language models can acquire decision-making skills through embodied interaction.
Read paper ↗2025
PolySim: Bridging the Sim-to-Real Gap for Humanoid Control via Multi-Simulator Dynamics Randomization
A whole-body control training platform that jointly trains policies across heterogeneous simulators, reducing simulator bias and enabling zero-shot deployment on a real humanoid robot.
Read paper ↗2024
Interruption-Aware Cooperative Perception for V2X Communication-Aided Autonomous Driving
A communication-adaptive spatiotemporal recovery framework that uses historical information, knowledge distillation, and curriculum learning to restore cooperative perception under V2X interruptions.
Read paper ↗2024
Robust Collaborative Perception without External Localization and Clock Devices
FreeAlign estimates relative pose and timing from shared object geometry, enabling robust collaborative perception without relying on external localization or clock signals.
Read paper ↗2024
Self-Localized Collaborative Perception
CoBEVGlue matches co-visible objects to recover relative poses, making collaborative perception robust to localization noise and adversarial pose errors.
Read paper ↗2023
Emergent Communication in Interactive Sketch Question Answering
An interactive sketch question-answering task and multi-round emergent communication system balancing answer accuracy, drawing complexity, and human interpretability.
Read paper ↗2022
Where2comm: Communication-Efficient Collaborative Perception via Spatial Confidence Maps
Spatial confidence maps identify perceptually critical regions to share, preserving multi-agent perception performance while dramatically reducing communication.
Read paper ↗2022
Latency-Aware Collaborative Perception
SyncNet compensates asynchronous features through attention-based estimation and time modulation, improving collaborative perception under communication latency.
Read paper ↗2022
Robust Collaborative Perception against Communication Interruption
A study of communication disruptions in collaborative perception that recovers missing information from historical observations to maintain robust detection performance.
Read paper ↗Technical Reports
No technical reports in this research direction.
2026
Qwen-RobotNav Technical Report: A Scalable Navigation Model Designed for an Agentic Navigation System
A scalable navigation model with a reconfigurable observation strategy, allowing a high-level planner to switch task modes and context strategies during long-horizon execution.
Read paper ↗2026
Qwen-RobotManip Technical Report: Alignment Unlocks Scale for Robotic Manipulation Foundation Models
A generalizable vision-language-action foundation model that aligns heterogeneous manipulation data across representation, motion, and behavior to enable large-scale training and cross-embodiment transfer.
Read paper ↗2026
Qwen-RobotWorld Technical Report: Unifying Embodied World Modeling through Language-Conditioned Video Generation
A language-conditioned video world model that predicts physically grounded futures across manipulation, autonomous driving, navigation, and human-to-robot transfer.
Read paper ↗2026
Qwen-VLA: Unifying Vision-Language-Action Modeling across Tasks, Environments, and Robot Embodiments
A unified embodied foundation model for manipulation, navigation, and trajectory prediction, extending vision-language reasoning into continuous action generation across robot embodiments.
Read paper ↗2025
Bohrium + SciMaster: Building the Infrastructure and Ecosystem for Agentic Science at Scale
An infrastructure-and-orchestration stack for traceable, reproducible long-horizon scientific workflows, turning diverse scientific assets into reusable agent-ready capabilities.
Read paper ↗Experience
Education and industry.
Education
-
Shanghai Jiao Tong University
Ph.D., School of Artificial Intelligence
2024 – 2027expected
-
Shanghai Jiao Tong University
M.Eng., Cooperative Medianet Innovation Center
2021 – 2024
-
Shanghai Jiao Tong University
B.Eng., Electronic Engineering · Outstanding Engineers Class
2017 – 2021
Industry
-
Tongyi Qwen
Core contributor of Qwen-Robot Suite (Manip and Nav)
2026 – Present
-
ByteDance
Multimodal LLM Engineer, DATA
2024
-
Intel
Graphics Software Engineering Intern, IAGS
2020 – 2021