Yitong Hu

Intelligence,
in a human world.

I’m Yitong Hu 胡逸同, a master’s student at BUPT. I’m exploring how intelligent systems perceive, act, and collaborate in the world we share.

Explore my work Curriculum vitae
Different perspectives, a shared worldTwo differently oriented sets of nested ellipses represent the perspectives of a system and a human. Their overlap is softly shaded. A gently winding path passes through both participant centers. The system has a nine-dot matrix, with dots becoming smaller from the center outward like a quiet ripple; the human has a single-line head profile facing the system. This is a conceptual drawing.
Currently exploring

Embodied intelligence Human–agent interaction Collaborative systems

01 / Selected work

Ideas, put into practice.

2026 Embodied AI · Human–AI teamwork

TeamHab

A platform for embodied human–AI teamwork.

Two people wearing VR headsets and two simulated Spot robots, Yellow and Blue, taking part in a shared embodied task.
Two people, two robot agents, one shared environment

What does it take for people and robots to work as one team? TeamHab supports scalable teams of VR participants and AI-controlled robots working in shared Habitat environments. Robot planning and team communication are configurable, while views, messages, actions, and outcomes are recorded together. The demo follows a two-person, two-robot team as they coordinate and hand over parts of a task.

Research prototype · Work in progress

Video

2026 Embodied AI · Social navigation

Listen to Yield

Hearing beyond the visible.

An illustrated indoor navigation scene showing how a robot uses human calls, footsteps, and conversation to navigate around people.
Socially-aware audio-visual navigation

How can a robot make room for someone it cannot yet see? SAVNav turns sounds such as footsteps and conversation into spatial constraints. Its audio-visual map anticipates people beyond the robot’s line of sight, guiding active exploration and socially compliant navigation. The training-free framework is evaluated in Habitat and deployed on a physical Stretch 3 robot.

IEEE RA-L · Under review

ManuscriptVideo

2026 LLM agents · Context evolution

Unified Context Evolution for LLM Agents

Turning experience into reusable context.

A comparison of ReAct and UCE, showing how a library of memory, strategy, workflow, and skill feeds reusable context into an LLM agent.
From fixed context to an evolving context library

What should an agent carry from one task to the next? UCE organizes experience into an evolving library of memory, strategy, workflow, and skill. These context units are retrieved during decisions, evaluated through use, and refined over time. The framework improves performance on interactive tasks without retraining the underlying model.

EMNLP 2026 Main Conference · Accepted

Paper

2025 Multi-robot systems · Adaptive teamwork

Proactive Collaboration

Robot teams that adapt their membership.

Three rows compare fixed, responsive, and proactive robot collaboration, showing how anticipating a capability gap allows a team to recruit the member it needs.
Fixed, responsive, and proactive collaboration

Can a robot team recognize when it needs a different set of teammates? This work studies proactive collaboration through autonomous interaction. Robots discuss task progress, negotiate responsibilities, and recruit or release members as needs change. Experiments in simulation and on physical robots explore how anticipating capability gaps can support more adaptive teamwork.

Nature Communications · 2026

PaperCode

2024 Human–agent interaction · Driving

HCockpit

A shared understanding of the road.

HCockpit workflow showing an in-car view, a copilot agent, and a bird's-eye view from a driving simulation.
Human–agent collaboration in a simulated cockpit

What does an AI copilot need to understand about the driver as well as the road? HCockpit brings together information from inside and outside a vehicle to explore context-aware coordination between people and agents.

Undergraduate project

ReportCode

2023 Information systems · Research discovery

CollabXplorer

Connections across disciplines.

CollabXplorer dashboard showing an academic profile, suggested research collaborators, and a colored network of scholarly connections.
Exploring research through scholarly connections

How might we find collaborators beyond familiar academic circles? CollabXplorer combines research semantics and scholarly networks to support the discovery of potential interdisciplinary collaborators.

National-level undergraduate innovation project

ReportCode

02 / A little background

Curiosity across
boundaries.

My work has taken me from discovering connections between researchers to exploring collaboration between people, agents, and robots.

I’m interested in how perception and context can inform intelligent behavior—and how that behavior fits into people’s lives. These are questions I’m continuing to explore.

More in my CV

  1. 2024 — present

    Beijing University of Posts and Telecommunications

    Master’s student · AI

  2. 2023 — 2025

    Tsinghua University

    Visiting research assistant

  3. 2023

    Peking University

    Research intern

  4. 2020 — 2024

    BUPT & Queen Mary University of London

    Undergraduate studies

03 / In conversation

Good work starts
with a conversation.