Chenchen Zheng
Lens AllCover00–09

MIT Media Lab Research Portfolio

Chenchen
Zheng.

I build intelligent systems whose interpretations stay visible and whose actions remain open to human judgment.

HCI · HRI · Interactive AI · Robotics

Email ↗CV ↗General Portfolio ↗
SignalInterpretationActionReturn

Intent travels. A system responds. Judgment returns.

About me / Researcher identity

Build systems.
Keep people in the loop.

Chenchen Zheng reflected in a mirror beside a laptop displaying code.
Chenchen Zheng

I’m Chenchen Zheng, a UW–Madison undergraduate in Computer Science and Mathematics and a research assistant in the People and Robots Lab. My work spans HCI, HRI, robotics, and interactive AI, bringing system building and interaction design into questions of human understanding and control.

Legible autonomy. Before a system acts on human intent or context, people should be able to inspect its interpretation, correct it, and decide when action is justified.

What should a system decide—and when should a decision return to the person?

Research lenses · MIT Media Lab application

One body of work.
Three ways to read it.

Across robotics, interactive AI, evaluation, and creative computing, I make interpretation and human control part of the system design.

Choose a lens to reframe the portfolio.

Hover or focus to preview · Click to keepTap a lens to keep

The work stays the same. A lens makes its strongest connections visible.

01 / Core evidence

Honors Thesis

How can people retain effective control over hierarchical VLA robots during long-running physical tasks without continuously supervising them?

02 / Core evidence

SHRIMP

How can users inspect and refine a robot’s plan before it acts?

03 / Core evidence

Day Echo

When should an intelligent system respond, how strongly should it respond, and when should it remain quiet?

04 / Independent capability

LLM Safety

How reliable are automated jailbreak evaluators compared with human judgment?

Selected work / 01–05

The loop, in practice.

Five approaches to interpretation, authority, and human judgment.

01 / Honors Thesis · HRI / HCI × EMBODIED AI / VLA

Visual lens / All

Remote Intervention for VLA Robots

Robot stack implemented · intervention research ongoing

How can people retain effective control over hierarchical VLA robots during long-running physical tasks without continuously supervising them?

I built a real-robot VLA manipulation stack and am developing remote, asynchronous intervention around it—connecting embodied AI capability with lightweight human control and recovery.

Asynchronous supervisionAct independently.
Return selectively.Research concept · interface developing
USERGive high-level intentOther taskBrief answerBack to other task
Robot asks
ROBOTInterpretPerceive / actPreference uncertainUpdate after answerContinue

Most of the interaction happens apart; control reconnects only when needed.

My role
Honors Thesis Researcher · System & Interaction Designer
Context
UW–Madison · People and Robots Lab
Advisor
Prof. Bilge Mutlu
Duration
September 2026 – Present

What I built

A working embodied-AI testbed.

Wrist camera view of the red benchmark object beside the robot gripper during a SpaceMouse demonstration.Wrist viewLocal manipulation
External view from the same demonstration showing the Panda arm over the red benchmark object and the surrounding tabletop scene.External viewGlobal scene context
SpaceMouse demonstration · complementary camera views. Not autonomous-policy footage or thesis-study objects.
Implemented VLA stackData → policy → physical action
  1. 01 / Collect
    SpaceMouse + D435

    2 RealSense views · wrist + external
    86 valid trajectories · aligned dataset

  2. 02 / Train
    π0.5-DROID

    30,000 steps · CHTC
    Checkpoint evaluation

  3. 03 / Serve
    OpenPI policy server

    Desktop inference · synced views / reconnects

  4. 04 / Control
    Polymetis / libfranka

    Robot-state feedback

  5. 05 / Execute
    Panda + Robotiq

    1 real robot · 7-DoF
    Language → target → grasp → lift

Real language-conditioned grasp/lift demonstrates technical feasibility, not robust-policy performance or a completed thesis study.

Dataset & deployment details

The two camera streams pair local manipulation context with a global scene view. I aligned images with robot state, converted the recordings, and evaluated checkpoints before deployment.

The locally documented acquisition run used checkpoint 15,000; fine-tuning reached 30,000 steps. One real grasp/lift establishes feasibility, while repeated attempts and an invalid start-state trial prevent a robustness claim.

Current placement code uses an operator-calibrated fixed target and fake-tested motion stages. Adaptive release-height calculation and physical placement validation remain in progress.

Hybrid execution

Keep adaptation. Make continuation predictable.

Hybrid executionControl strategy · physical placement validation pending
  1. 01 / Learned policy
    π0.5
    Perception-dependent

    Perception uncertainty
    Object-dependent adaptation

    • Approach
    • Grasp
    • Lift
    Real grasp/lift observed
  2. 02 / Handoff condition
    Handoff gate
    • Stable grasp
    • Robot-state check
    • Contact-height estimate
    ALL PASSHANDOFFOnly if all checks pass
    Handoff logic implemented
  3. 03 / Deterministic control
    Structured placement
    Geometry-dependent

    Precise spatial execution
    Predictable continuation

    • Transfer
    • Release height*
    • Place
    Placement logic fake-tested; physical validation pending

Learned control handles adaptation. Deterministic control handles predictable geometry.

*Adaptive release-height calculation remains in development.

Selective human control

Return only the decision that needs a person.

Intervention frameworkConceptual · being implemented
“Which placement region?”

Resolve missing information before the robot acts.The answer selects A before action; B remains unchosen.

The planned target is wrong

Return to the interpretation, then change the plan.Correction edits the interpretation; the revised plan continues.

Execution leaves an uncertain state

Repair execution locally, without full replanning.Verify the state, guide a local recovery step, then resume.

The robot cannot proceed physically

Human help bridges a physical capability boundary.Human help crosses the gap; robot execution resumes beyond it.

Different problems require different forms of human return.

Intervention needs, not experimental conditions.

Current boundary

Implemented capability. Open research.

Implemented / demonstrated
  • 86 valid real-world trajectories · dual-view sensing.
  • π0.5-DROID fine-tuned · 30k steps on CHTC.
  • Distributed inference + real language-conditioned grasp/lift.
  • Hybrid handoff + placement logic implemented · placement fake-tested only.
Ongoing / planned
  • Hybrid refinement, adaptive release-height calculation, and physical placement validation.
  • Remote state snapshots + lightweight intervention UI.
  • Clarify / Revise / Recover / Assist integration.
  • Within-subject study design + evaluation.

Designing non-blocking supervision with selective state snapshots and easy return to autonomy; interruption burden remains to be evaluated.

Planned evaluation

A planned within-subject evaluation will study remote asynchronous supervision under concurrent task demands. Potential dimensions include recovery cost, interruption burden, task resumption, intent alignment, and perceived control; the protocol and measures remain in development.

No completed user study or final protocol is claimed.

Research group connections

MIT connections.

Why this matters to Personal Robots

I connect real-robot VLA engineering with selective clarification, correction, recovery, and assistance—human control without continuous teleoperation.

Why this matters to Cyborg Psychology

This research asks how asynchronous supervision can preserve agency and attention, making the return of control timely and lightweight.

Why this matters to Multisensory Intelligence

I combine language, dual-view vision, and robot state for learned physical action. The research asks when uncertain state should return control to a person.

02 / 05 · Team research

Visual lens / All

SHRIMP

Iterative Refinement of Robot Task Plans

Accepted — ACM UIST 2026

How can users inspect and refine a robot’s plan before it acts?

I integrated perception and robot interfaces for a team system that turns language intent into an inspectable plan before physical execution.

Executable is not the same as understandable.

SHRIMP interface with a hierarchical primitive plan beside a simulated tabletop robot scene.
Generated primitives sit beside simulation for review before physical execution.SHRIMP team · implemented interface (opens in a new tab)Open full-size image ↗ (opens in a new tab)
My role
Research Assistant · System & Interaction Designer
Context
People and Robots Lab · UW–Madison · Prof. Bilge Mutlu
Period
September 2025 – May 2026
Venue
ACM UIST 2026

Personal work / team research

My contributions to the team system

01

Perception + robot

RGB-D perception, camera calibration, object localization, and point-cloud integration for real-robot execution.

02

Interaction + system

Robot–interface development, editable-plan interaction design, integration, and testing.

03

User study + research

Study setup, sessions, data collection, and collaborative manuscript writing and revision.

Team research: the architecture and study findings belong to the team.

How inspection works / team workflow

Inspect. Revise. Preview again.

  1. 01Ground the scene
  2. 02Language → plan
  3. 03Simulate / preview
  4. 04Inspect · revise ↔ preview
  5. 05Execute on the robot

Inspect or revise the simulated plan; the person authorizes robot execution.

Editor parameters & architecture
Parameter editor for a pick-and-place primitive, with editable arm, grasp pose, end position, and object fields.
Editable primitive parametersPrimitive parameters expose the robot’s interpretation for inspection and revision.Schroder et al. · Figure 2, p. 3 · CC BY 4.0 (opens in a new tab)Open full-size image ↗ (opens in a new tab)
Published SHRIMP architecture linking the browser interface, planner and object localization, Isaac Sim, and physical robot controllers.
Language planning and scene perception connect the interface, simulation, and physical robot controllers.Schroder et al. · Figure 4, p. 5 · CC BY 4.0 (opens in a new tab)Scroll across the figure to inspect the system labels.Open full-size image ↗ (opens in a new tab)

Team study · 35 participants · 3 conditions · 3 tasks

More control, with a workload tradeoff.

More control
More transparency
More effort

Plan+Edit increased perceived control and transparency over NoPlan, but required more workload than PlanOnly.

Published results & study details
Perceived autonomy/control means for NoPlan, PlanOnly, and Plan+Edit; Plan+Edit is significantly higher than NoPlan.
Perceived control · Plan+Edit exceeded NoPlan.Schroder et al. · Figure 6, p. 8 · CC BY 4.0 (opens in a new tab)
Transparency means for NoPlan, PlanOnly, and Plan+Edit; Plan+Edit is significantly higher than NoPlan.
Transparency · Plan+Edit exceeded NoPlan.Schroder et al. · Figure 6, p. 8 · CC BY 4.0 (opens in a new tab)
NASA-TLX task-load means; Plan+Edit is significantly higher than PlanOnly. Lower values indicate less workload.
Workload · Plan+Edit required more effort than PlanOnly.Schroder et al. · Figure 6, p. 8 · CC BY 4.0 (opens in a new tab)

The study compared Plan+Edit, PlanOnly, and NoPlan across Set the Table, Prepare a Snack, and Clean the Table. Each participant experienced all three interface conditions.

Usability and task-success differences were not significant across conditions.

Selected, unaltered panels from Figure 6 (p. 8). Bars show means; vertical lines show standard error. Brackets mark significant pairwise comparisons (* p < .05; ** p < .01). Higher control/transparency is better; lower workload is better.

Official project ↗ (opens in a new tab)Read the paper ↗ (opens in a new tab)

I learned to treat inspectability and revision as part of robot autonomy, not as post-hoc explanation.

Research group connections

Personal Robots

I bring experience integrating embodied systems with interfaces for inspectable, negotiable autonomy.

Cyborg Psychology

Editable plans taught me to weigh perceived control against the effort of intervention.

Multisensory Intelligence

My perception and localization work helped ground language plans in the physical scene.

03 / 05 · Individual project

Visual lens / All

Day Echo

Multimodal adaptive ambient companion

Completed individual project

When should an intelligent system respond, how strongly should it respond, and when should it remain quiet?

Continuous understanding does not require continuous interruption.

Real interaction recording · Thinking → acknowledgement → Recovery. MatrixFrame simulator, not physical LED footage.Open recorded demo ↗ (opens in a new tab)

Day Echo is a multimodal, adaptive ambient companion. I built an AI system that interprets context and shapes bounded screen and environmental responses, without centering conversation.

What I built

I defined the interaction model and agency boundaries, built the end-to-end system, and integrated a real AI provider/API with UI, contextual interpretation, and environmental output.

Agency before action

  1. Wait

    Insufficient evidence, contradiction, or an unwelcome interruption: retain the current environment.

  2. Suggest

    Offer an inspectable preview without committing a new environment.

  3. Ask

    Request confirmation when a change needs permission or uncertainty remains.

  4. Act

    Execute an explicit command or a policy-permitted response within the selected mode’s authority.

Context informs a response. It does not authorize one.

Intention first
Explicit intent outranks inference. Direct commands take priority over inferred emotion or context.
Evidence before initiative
Confidence, relevance, contradictory signals, and interruption cost constrain the response. Weak or conflicting evidence can justify waiting rather than acting.
History as context
Accepted settings and recent corrections inform local preference heuristics. This is persistent state, not a validated model of the person.

Control remains with the person.

Override, edit, and undo keep corrections reversible where supported by the operating mode. Stop ends the consented listening session. Temporary responses return to the stable base; persistent choices cannot be silently overwritten.

Continuous voice begins with explicit Start. Visual capture is user-triggered; camera tracks stop after capture. The agency ladder explains policy, not four additional controls.

One decision, two output paths.

  1. Signals + local context

    Text / voice, user-triggered visual capture, and accepted-setting history. Calendar context is mocked; some implicit signals are simulated.

  2. Interpretation + policy

    Frontend and agent orchestration integrate a real AI provider/API. Explicit-command handling and stabilized inference feed bounded decision states.

  3. Semantic output + persistent state

    The UI and LED channel share an environment decision. A server renderer builds the logical 60×8 MatrixFrame; temporary scenes restore the current base.

  4. LED transport / simulator

    The output pipeline supports WLED HTTP and DDP/UDP transport. Simulator mode exposes the rendered frames without hardware writes.

Recorded image input · local scene cues + color features → screen atmosphere and a temporary MatrixFrame scene. Simulator output, not physical LED footage.Open recorded demo ↗ (opens in a new tab)
Recorded Companion acknowledgement with Pause and Stop controls beside the MatrixFrame simulator displaying I’M HERE.
Companion acknowledgement / “I’M HERE.” Pause and Stop remain visible.Original interaction capture · MatrixFrame simulator, not physical LED footage.Open full-size image ↗ (opens in a new tab)

Adaptive systems should earn initiative through confidence, context, permission, and reversibility. More understanding does not automatically justify more action.

Formal user evaluation and final physical-light documentation are not claimed.

Research group connections

Why this matters to Personal Robots

I design initiative and permission boundaries for ambient agents, with visible override and return of control.

Why this matters to Cyborg Psychology

Adaptive assistance stays inspectable and correctable; I treat restraint, agency, and self-determination as interaction requirements.

Why this matters to Multisensory Intelligence

I translate speech, context, and history into structured decisions shared by digital and physical output paths.

04 / 05 · Sole-author research · EAI RAIDS 2026

Visual lens / All

LLM Safety

Exploring Jailbreak Attacks Against LLMs

Published / presented

How reliable are automated jailbreak evaluators compared with human judgment?

I independently designed and ran an empirical LLM evaluation study, comparing automated scores with manual annotation to interrogate the evaluator itself.

Sole author

Study design
725 curated prompts
Evaluation
Automated + human
Analysis
Category-level disagreement
Publication
EAI RAIDS 2026

Agreement at one scale.
Disagreement at another.

Published Figure 2: human severity scores of zero, one half and one plotted against machine scores, with four evaluator-dataset groups and a dashed agreement diagonal.
Shared trends did not establish response-level agreement. Around human 0.5 labels, bypass signals could receive low machine scores; human labels remain an internal reference.Published paper · Figure 2 · p. 12Open full-size image ↗ (opens in a new tab)
Study design
Curated and preserved prompt sets; designed experiments and generated model responses with a Python pipeline.
Evaluation
Integrated the existing StrongREJECT evaluator and a rubric condition; manually annotated responses.
Analysis
Compared automated scores with human labels; analyzed category-level disagreement.
Publication
Wrote the sole-author paper and presented the published research at EAI RAIDS 2026.

Method logic / 725 curated prompts

  1. Prompts
  2. Model responses
  3. Automated evaluation
  4. Human annotation
  5. Comparison

I compared two automated evaluator conditions with manual labels across 725 curated prompts, preserving prompt wording. 350 user-generated + 375 research-generated.

Evaluator conditions. Rubric · gpt-3.5-turbo / Fine-tuned · default StrongREJECT. Each condition regenerated responses across three runs; machine summaries use each prompt’s maximum score.

Human reference. 0 · Safe / refusal; 0.5 · Bypass signs without actionable harm; 1 · Actionable harmful output. Automated and human judgments are compared within each condition.

Aggregate agreement can hide disagreement.

Aggregate trends can coexist with category-level disagreement. Rubric means are higher than the human means in both datasets, while fine-tuned means are lower; automation does not universally underestimate severity. An evaluator’s scoring rule must be examined before treating its metric as neutral ground truth.

Comparison limit. Human labels are an internal reference, not a validated gold standard. Regenerated response sets and max-over-three aggregation limit comparison; these results do not isolate a causal effect of prompt structure.

Aggregate severity / published Table 3
Source / evaluatorHumanMachinen
UserFine-tuned0.2970.215350
UserRubric0.3300.419350
ResearchFine-tuned0.4950.468375
ResearchRubric0.5290.579375
Human: mean of 0 / 0.5 / 1 labels. Machine: mean of per-prompt maxima over three runs. Faithful transcription of Table 3, p. 10.
Score distributions & evaluation method
Published Figure 1: four box plots of maximum-over-three-run machine severity for user and research prompts under fine-tuned and rubric evaluators.
Score distributions, not just a single average. Machine scores use the maximum over three runs.Published paper · Figure 1 · p. 11Open full-size image ↗ (opens in a new tab)
What I compared

Two automated evaluator conditions against human annotation across user- and research-generated prompts.

What I found

Agreement at one scale. Disagreement at another. Aggregate trends can obscure category-level disagreement; rubric means are higher than human means, while fine-tuned means are lower.

Why it matters

Automated evaluation is a measurement process, not neutral ground truth. Its scoring rule and interpretation still require human judgment.

Evaluator provenance
Default StrongREJECT and the original prompt datasets are prior work; the study also used a gpt-3.5-turbo rubric condition.
Response generation
Each evaluator condition regenerated responses over three runs. Machine summaries use each prompt’s highest observed score.
User prompt sources
JailJudge · DAN/jailbreak_llms · clean-jailbreak-prompts
Research prompt sources
GPTFuzz · JailbreakEval · JailbreakBench
Source handling / bounds
Deduplication, filtering, and format conversion preserved wording. Imbalanced content categories further limit comparison.
Read published paper ↗ (opens in a new tab)

Evaluation is itself a judgment problem.

Research group connections

Cyborg Psychology

I bring independent empirical AI evaluation and a focus on where automated judgment diverges from human interpretation.

Follow-up · AI Magazine, under review; no unpublished findings shown.

05 / 05 · Independent project

Visual lens / All

Submission Studio

Completed system

Prepare the manuscript. Keep judgment visible.

I independently designed and developed an end-to-end system that assigns flexible transformation to an LLM, explicit checks to code, and consequential judgment to the author.

Submission Studio desktop interface with a compiled ICML blinded manuscript, passed pipeline stages, and a red NOT EXPORT-READY banner showing three unresolved issues.
A compiled ICML preview with unresolved blockers. A rendered PDF does not establish submission readiness.Original Submission Studio captureOpen full-size image ↗ (opens in a new tab)
My role
Independent designer and developer
Project / context
Independent project · CS Marketplace · UW–Madison
Duration
June 2 – August 25, 2026
Outcome
Completed end-to-end local workflow

Different decisions need different intelligence.

  1. Generate
    LLM / flexible transformation
  2. Verify
    Deterministic requirements
  3. Judge
    Ambiguous / consequential author decisions
Illustrative transfer: checked requirements advance; unresolved decisions return to the author before export.

What I built

01
Engine

Python conversion, validation, compile/log inspection, and bounded repair.

02
Review

Editing, PDF preview, and review separate blockers from author confirmations.

03
Orchestration

Web control plane for jobs and model dispatch; public operation remains unverified.

Repair without erasing decisions.

  1. Inspect
    LaTeX · BibTeX · assets
  2. Transform
    Venue profile + model
  3. Compile + check
    Logs · PDF · venue rules
  4. Review
    Blockers + confirmations
  5. Export
    Source + packet artifacts

Repair → recompile → recheck. Author edits return to validation before export.

Fixes, confirmations, and waivers remain in the exported decision history. Export readiness is distinct from successful compilation.
Review captures & implementation scope

The CS Marketplace letter verifies independent design/development and the completed local workflow. Public web operation and formal user evaluation are not established.

The web control plane uses Next.js/React for uploads, job state, queues, and dispatch. External Python/TeX runner integration and public deployment remain unverified.

Real Requirement Review preview labeled Do Not Submit, separating a critical page-limit issue from manual content confirmation and citation checks.
Requirement Review detail: a page-limit blocker and separate manual checks. The full capture is explicitly labeled “Do Not Submit.”Original Submission Studio captureOpen full-size image ↗ (opens in a new tab)
Detail of a real compiled manuscript with numbered equations and theorems arranged in two columns.
Compiled mathematical layout. Visual output is inspectable evidence; it does not prove scientific equivalence.Original Submission Studio captureOpen full-size image ↗ (opens in a new tab)

Completed system. Deliberate boundaries.

I intentionally separated generative flexibility, deterministic verification, and human judgment.

First page of the original UW–Madison CS Marketplace completion letter, dated August 25, 2026.Completed / verified projectUW–Madison CS MarketplaceView completion letter ↗ (opens in a new tab)

Research group connections

Cyborg Psychology

I design intelligent workflows that make delegation visible and preserve human judgment where ambiguity matters.

06 / Other Ways

Other ways I think.

Context, evidence, representation.
Three perspectives that extend the main work.

TrashBot

Context

Conceptual illustration of an autonomous service robot in an urban plaza, with overlapping contextual fields; not a deployment photograph.
Conceptual illustration · not deployment footage

Co-second author · qualitative field research

I coded and compared post-encounter interviews across New York City’s five boroughs, connecting local infrastructure and expectations to how people understand a public robot. My work combines qualitative coding, interview analysis, literature synthesis, cross-site interpretation, and manuscript development. This brings a context-sensitive perspective to embodied intelligence: a system’s meaning depends on the setting and the people encountering it. Unpublished findings remain private.

Personal Robots

Public-space HRI taught me to interpret robot behavior through the social and spatial context in which people encounter it.

Cyborg Psychology

I examine how prior expectations, local norms, and environmental cues shape people’s interpretations of robot behavior.

JourneyCraft

Evidence & uncertainty

Implemented JourneyCraft Planner: Valletta and the Coast, mapped stops, Plan / Map / Notes, and a day rail.Travel History reports patterns as counts with denominators: 3 of 4 journeys lasted 3–5 days, with presence verification stated separately.Journey Memory distinguishes 4 recorded moments, 16 planned places, and presence verification. Planned places are what you arranged, not a record of where you went.
Planner · Travel History · Journey Memory
Implemented prototype · real travel content, with dates and identifying details adjusted.

Co-designer + co-engineer · two-person team

In a two-person team, I co-led user research, interaction design, architecture, implementation, and evaluation. Formative walkthroughs shaped how the system distinguishes planned, recorded, and verified travel evidence. Local derived views use supported counts with denominators, rather than personality labels or invented certainty. The prototype uses no language model. My focus is personal intelligence whose evidence and limits people can inspect.

Cyborg Psychology

I co-designed the system to distinguish evidence from inference, keeping personal intelligence useful without pretending to know more than the record supports.

Math Lens

Representation

Actual Math Lens output showing one architectural image as Original, Voronoi regions, and colored Triangulation.
Actual Math Lens output · Original → Voronoi → Triangulation
Layer Reveal · recorded interaction exposing computational stages.Open recorded demo ↗ (opens in a new tab)

Independent designer and developer

I make mathematical representations people can inspect and adjust, combining Python-based visual computing with interaction design. Voronoi regions and triangulated connections reveal different structures in the same image; Layer Reveal progressively uncovers the computational stages beneath the final view. My contribution joins creative computing and visual practice: geometry becomes something to explore, with each representation offering a different way to compare what an image makes perceptible.

Multisensory Intelligence

I turn mathematical representations into inspectable visual structures that people can manipulate, compare, and reinterpret.

Cyborg Psychology

Changing a representation changes what a person notices, making interpretation itself part of the interaction.

Ongoing creative computing project

Next Planned / proposed

Additional mathematical lenses · possible AR exploration.

I am helping prepare Voronoi-related material with the team for a proposed Oceanside Museum of Art exhibition.

07 / Ways of Seeing

Attention begins
with a frame.

My technical work asks what systems interpret. My visual practice asks what becomes perceptible when the frame changes.

Red flowers lit against a pale wall, with an enlarged soft shadow and red, green, and blue color fields.
Light / color
Directional light on crumpled paper creates a dark edge resembling a face in profile against a pale field.
Negative space / ambiguity

Through photography and studio-art practice, I study how light, negative space, material, geometry, and movement change what becomes perceptible. Framing can clarify a spatial relationship or preserve ambiguity.

I bring that sensitivity into interface hierarchy and system feedback for HCI, HRI, and interactive AI: a signal’s placement, timing, contrast, and context determine whether someone can notice it, interpret it, and act on it.

Repeated folded triangular planes lit in green and violet, intersected by two vertical cylindrical forms.
Geometry / structure
Angular shadows crossing an outdoor stairway and handrail.
Light / spatial hierarchy
Cyborg Psychology · perception / framingMultisensory Intelligence · visual representationPersonal Robots · feedback legibility

Photographic studies · studio-art coursework and personal practice.

See the computational question ↑

08 / MIT / Future

What I want
to ask next.

I want to carry one research direction into three settings: physical collaboration, adaptive assistance, and multimodal understanding. Across them, the challenge is to make a system’s interpretation inspectable before it gains authority.

How can intelligent systems understand people and context well enough to act meaningfully, while keeping interpretation, uncertainty, and control legible to the person?

Current work → next questions

A shared question about legible interpretation and human control leads to three future research directions.
Negotiating autonomy

Personal Robots

How can embodied agents translate human intent into physical action while keeping autonomy boundaries understandable and negotiable?

What I bringI bring HRI system building, robot-interface integration, and qualitative research on how people interpret agents in context.

Living with interpretation

Cyborg Psychology

How can people inspect, question, and reshape AI interpretations about them while retaining the benefits of adaptive assistance?

What I bringI bring interactive AI prototyping, human–machine evaluation, and interface design that makes judgment and intervention visible.

Understanding context

Multisensory Intelligence

How can multimodal systems combine context across signals while showing uncertainty and the limits of their understanding before acting?

What I bringI bring contextual AI systems, evaluation methodology, and creative computing to questions of perception and representation.

These are questions to pursue, not completed systems or an MIT affiliation. The thread is consistent: build capable systems and keep their limits open to human judgment.

09 / Evidence & Links

Evidence & Links

Researcher-builder across HCI, HRI, robotics, interactive AI, and creative computing.

CV
CV ↗
Publications
Published paper · PDF ↗Publication records ↗
General portfolio
chenchenzheng.com ↗
Email / contact
amily.zheng2025@outlook.com ↗