Human-Robot Interaction & Designing for Machine Users

Part 1 covers human-robot interaction (HRI) — a mature research field (its own ACM/IEEE conference since 2006) about humans sharing physical space with embodied machines. Part 2 inverts the lens — AI agents as users of interfaces: established science vs. emerging practice.

Part 1 — Human-robot interaction

Motion is communication

Legible ≠ predictable (Dragan, Lee & Srinivasa, HRI 2013): legible motion lets an observer infer the goal early ("it's reaching for the red cup"); predictable motion matches expectations for a known goal. They can conflict — legibility often means exaggerating the trajectory toward the goal, like an animator's anticipation pose. Design motion to telegraph intent, not just minimize cost.
Cue before acting: orient before moving, decelerate on approach, signal turns with light animations (a delivery-robot/AMR convention) — unannounced motion reads as unsafe even when it is safe.

Gaze is a status display: people read where a robot "looks" as what it will do next. Rethink Robotics' Baxter (2012) set the exemplar: its screen-face eyes glanced toward where the arm would reach next — a safety instrument, not decoration.
Proxemics transfer to robots. Hall's zones (intimate <~45 cm, personal ~0.45–1.2 m, social 1.2–3.6 m) apply: people distance themselves from robots much as from people, modulated by robot gaze and likability (Mumm & Mutlu, HRI 2011). Respect personal space, approach from the front within view, slow as distance closes.
The uncanny valley (Mori 1970; authorized translation Mori, MacDorman & Kageki 2012, IEEE RAM): affinity rises with humanlikeness then plunges near-but-not-quite-human; motion amplifies it. Evidence is mixed in detail, but the caution holds: stylized beats almost-human.
Form is a promise: humanlike appearance inflates capability expectations and unearned social trust — the embodied version of the fake-anthropomorphism anti-pattern in Designing AI-Powered Interfaces. Eyes on a machine that can't see are a lie.

Cobots & industrial robots

Safety standards, current state: ISO 10218-1/-2 got a generational revision in 2025; most of ISO/TS 15066:2016 (the cobot spec: power/force limiting, speed & separation monitoring) was absorbed into ISO 10218-2:2025, and "collaborative robot" became "collaborative application" — collaboration is a property of the application, not of the robot.
Speed & separation monitoring is UX: the robot slows as a person nears, stops within protective distance, resumes when clear; make each state visible (lighting/HMI) so a slowdown reads as "it sees me," not "it's broken" — visibility builds shared-workspace trust.
Stop clarity and status lighting: e-stop is the red mushroom on yellow (ISO 13850); distinguish protective stop (auto-resumable) from e-stop (deliberate reset); restart must never surprise — announce before motion resumes. Stack lights follow IEC 60204-1 semantics (red fault, yellow abnormal, blue action required, green normal) — don't reassign those colors.

Consumer robots (vacuums, mowers, delivery)

"I'm stuck" is the core error UX: report location + cause + one recovery action in the app, plus an on-device sound so it can be found; fail safe (park, don't drain battery retrying); no blame, no fake distress; don't cry wolf — an over-notifying robot gets muted, then lost.
Household coexistence: pets and children treat robots as agents — design for being chased, ridden, and blocked; slow, predictable motion.
Onboard cameras are a privacy surface: disclose what's captured, where processed, who can view — including contractors (see the 2022 leak of Roomba development-unit images via data-labeling gig workers, MIT Technology Review). Shutters and recording lights beat policy text.
Sidewalk etiquette (exemplars: Starship — 10M+ deliveries by April 2026; Serve Robotics on Uber Eats in LA): yield to pedestrians, keep paths predictable, signal at crossings, stay visible (flag, lights), hand off to remote human help when confused rather than blocking the way.
Expressive nonverbal sound: R2-D2-style semantic-free utterances and earcons convey state and affect without promising language competence. Canonical case: Anki's Cozmo (2016) — character direction by ex-Pixar animator Carlos Baena, a film-style animation pipeline driving an "emotion engine," deliberately simplified eyes: appeal and legible state without humanlike over-promise (Anki folded in 2019).

Teleoperation

Show link latency/staleness explicitly — operators must know if they see now or 800 ms ago; fuse egocentric camera video with an allocentric map (camera alone loses spatial context); prefer supervisory control over joysticking as latency rises. Depth: Aviation, Marine & Drone Interfaces (ground-control-station section).

Humanoids (2025–26 wave) — emerging, unsettled

Shipped vs. demo, verified mid-2026: Figure 03 deployed in the tens of units commercially (BMW Spartanburg, from Jan 2026); Tesla Optimus in production but internal-only; Unitree G1 ships in the thousands, mostly to researchers; 1X's NEO home pilot delivers to early adopters (~$20k) with autonomy backstopped by human teleoperators. Treat demo reels as marketing until third-party verified.
Open HRI questions: bystander consent (a camera-laden robot records people who never opted in — and a remote teleoperator may be watching through it: disclose when); capability disclosure (autonomous or human-piloted right now?); humanlike form at maximum over-expectation.

Part 2 — Designing for machine users (EMERGING; thin peer-reviewed evidence)

AI agents now operate GUIs and consume content: Anthropic's computer use (API, late 2024) and Claude for Chrome (2025); OpenAI's Operator (Jan 2025, retired Aug 2025), then ChatGPT Agent (Jul 2025) and the Atlas browser's agent mode. Names churn; the durable fact is that a growing share of your "users" are software acting for a human.
Accessibility is machine operability. Agents parse the DOM and accessibility tree the way screen readers do: semantic HTML, real <button>s, ARIA roles, labeled inputs, and honest machine-readable state help agents exactly as they help assistive tech — the strongest new argument for 06-accessibility/ practices. Keep selectors and labels stable and meaningful; agents break on churned class-name soup.
Agent-hostile design is the new dark-pattern frontier: interfaces that mislead agents (hidden instructions, decoy buttons) — and agents falling for human-targeted dark patterns, auto-accepting what a person would decline. Both argue for Designing AI-Powered Interfaces's human-in-the-loop review patterns.
Machine-readable surfaces, by maturity: schema.org and OpenAPI — established; "the API is the agent's UI": clear operation names, descriptions, and error messages are agent affordances. MCP (Model Context Protocol, Anthropic, Nov 2024) — de facto tool-connection standard; donated Dec 2025 to the Agentic AI Foundation, a Linux Foundation directed fund co-founded by Anthropic, Block, and OpenAI. llms.txt — community proposal, real but limited adoption (per Information Architecture, don't restructure yet).
Agent identification vs. bot-blocking — open problem. CAPTCHAs can't distinguish a scraper from a user's legitimate delegated agent; blocking all automation now blocks customers. Web Bot Auth (IETF draft, atop RFC 9421 HTTP Message Signatures) lets agents identify themselves cryptographically — in production at Cloudflare/Anthropic/ OpenAI in 2026, but still a draft; consent, rate limits, payment: open.
Design stance: one interface, not an "agent mode." A single semantic, accessible, well-structured UI serves humans, assistive tech, and machine users at once (rhymes with Apple's no-separate-AI-mode stance in Designing AI-Powered Interfaces). Where GUI automation is the wrong tool, offer a real API or MCP server; forked agent-only experiences drift.

Sources

Dragan, A., Lee, K. & Srinivasa, S. (2013). "Legibility and Predictability of Robot Motion." Proc. HRI '13, 301–308.
Mumm, J. & Mutlu, B. (2011). "Human-Robot Proxemics." Proc. HRI '11; Hall, E. T. (1966). The Hidden Dimension.
Mori, M. (1970). "Bukimi no tani." Energy, 7(4); authorized translation Mori, MacDorman & Kageki (2012), IEEE RAM, 19(2).
ISO 10218-1/-2:2025 (absorbs most of ISO/TS 15066:2016); ISO 13850; IEC 60204-1; A3 "Updated ISO 10218 FAQ" (automate.org).
Fast Company / Engadget on Anki Cozmo's animation pipeline (2016); MIT Technology Review (Dec 2022), leaked Roomba images.
Starship Technologies press (10M deliveries, 2026); Serve Robotics.
Anthropic, "Donating the Model Context Protocol…" (Dec 2025); Linux Foundation AAIF announcement; modelcontextprotocol.io.
IETF draft-meunier-web-bot-auth-architecture; RFC 9421; Cloudflare Web Bot Auth docs; llms.txt — llmstxt.org (Howard, 2024; community proposal).
ACM/IEEE Conference on Human-Robot Interaction — the field's venue.