Intent, Accessibility, and Human Override

You press Cancel task. The screen turns green. The robot is still rolling. Which part of this interaction failed?

A useful robot must do more than find a valid route. People need to understand what it plans to do, intervene through controls they can use, and distinguish a received command from evidence about the machine. This lesson continues 9.2's distinction between an intended effect and an observed result.

Learning outcomes

By the end, you can:

• Distinguish a predictable path from motion that makes a goal easier to infer.
• Replace a color-only or voice-only control/status design with redundant, operable information.
• Trace a cancellation through request, acceptance, result and physical-state evidence.
• Evaluate whether a restart is permitted under an explicit supervisory policy.

Prerequisites: actions and feedback from module 4; state machines and timeouts from module 8; task evidence from 9.2. No robot, paid software or video is required. Time budget: 10 minutes concepts, 10 minutes trace, 20 minutes lab, 10 minutes retrieval and reflection.

Before / During / After the video

Before: Sketch a robot at a fork leading to stations A and B. What must a person know before walking near the fork? Now predict whether “Cancel accepted” proves the robot is stationary.

During: Follow the eight numbered scenes. In scene 4, separate the closed dispatch gate from the cart's deceleration. In scene 5, ask what each evidence card actually proves. Pause at scene 7 before reading the answer.

After: Explain two independent reasons the cart must not restart in scene 7. Propose a control a person can use without distinguishing red from green or hearing narration.

Watch the complete Humanoid Hub tutorial. Read the descriptive transcript instead if preferred; each scene includes its visible mechanism and full narration. The retrieval pause is not a timed assessment.

1. Intent is information, not decoration

For a delivery cart, destination B, an indicated route and an honest current status are more useful than a generic flashing light. Show the intended action early enough for the person to respond. If the destination changes, update the preview rather than leaving yesterday's arrow on screen.

Predictability: given the destination, is this the path an observer expects? Legibility: from the observed motion, can an observer infer the destination? These are different questions and can conflict. Neither a familiar path nor an expressive gesture proves a route is physically safe. The HRI research establishes the distinction; our station label and preview are an original design example, not a claim that one visual cue works for everyone. Dragan, Lee and Srinivasa

Retrieval pause: A straight initial segment could lead to A or B. Is it enough to know that the segment is short? Explain which question remains unanswered.

2. Make intervention usable in more than one way

Use a named Cancel task button with a stop symbol and visible keyboard focus. Show words such as Cancellation requested, Canceling, Held and Motion unknown. Color may reinforce the distinction but should not be the only carrier. Provide caption/text alternatives to speech. A browser status message should also be exposed to assistive technology without unexpectedly moving focus. W3C color guidance, keyboard guidance, status guidance

More channels do not automatically make an interface accessible. Labels, reach, timing, hearing, vision, motor needs and cognitive load vary. Ask representative users to attempt the actual intervention and explain the resulting state. Do not infer a disability from a person's appearance. This lesson tests browser controls and event semantics, not a physical accessibility certification.

3. Four events, four different meanings

The diagram separates protocol messages from motion evidence. Read the equivalent four-stage table below; no color interpretation or video is required.

ROS 2 distinguishes acceptance into CANCELING from terminal CANCELED. A cancel request can be rejected; feedback and results must be associated with the right goal. The separate dispatch gate and HELD state below belong to our teaching model, not the ROS goal-state enumeration. ROS 2 Actions

When the person cancels, the supervisor immediately refuses new task commands. The current motion can persist while the request travels and the robot decelerates. Never relabel “request sent” as “physically stopped.”

4. Worked event trace

Our fictional simulation uses maximum feedback age 0.5 s and a near-zero speed tolerance less than 0.01 model units/s. These are lab choices, not hardware settings. Initial setup assumes a known stationary cart and a checked scene before the first confirmation.

“Held” deliberately does not mean “allowed to restart.” The toy restart rule requires all five checks: HELD supervisor, matching CANCELED result, fresh zero-speed evidence, plan matching the current scene, and deliberate confirmation. A stale report can invalidate the held classification. If communication is lost, use Motion unknown and keep dispatch blocked; the UI does not magically stop the robot.

Calculate before deciding

At model time 3.0 s, the last zero-speed observation was at 1.0 s. Age = 3.0 − 1.0 = 2.0 s. Since 2.0 > 0.5, the report is stale. A CANCELED result cannot make it fresh. Without deliberate confirmation, two conditions are missing. Obtain current evidence and repeat the policy checks; do not merely turn the indicator green.

5. A software cancel is not an emergency-stop system

The browser simulation sends no commands to physical hardware. Its cancel control depends on ordinary software and messages. Do not wire it to a motor or substitute it for protective equipment. Real systems need application-specific risk assessment, engineered protection and qualified verification. An emergency stop is itself a complementary protective measure, not a guarantee that hazards cannot occur. The cited manufacturer documentation is an example of this distinction, not a reset or operating procedure for learners. Universal Robots manual

Misconceptions to catch

• “The shortest path explains the goal.” It may remain ambiguous at a fork.
• “Red is self-explanatory.” Color perception and conventions vary; give words and symbols.
• “Accepted means canceled.” Acceptance begins cancellation handling; it is not the terminal result.
• “Canceled means safe to approach.” Protocol status is not a physical safety assessment.
• “Releasing the button means resume.” Not under this lab's deliberate-restart policy.
• “A late result is useful regardless of task.” It can describe a different goal; check identity.

Apply and summarize

Use the simulation lab and paper alternative provided with this lesson. Predict a trace before running it. Explain the first failed condition using the event time, goal identity and observation, not only a color.

The transferable pattern is make intent understandable → make intervention usable → acknowledge honestly → verify state → restart deliberately. It is a supervisory design pattern, not a certified safety mechanism. Next, module 10 asks which robot capabilities should be learned from examples and how to evaluate them.

Further source details and applicability limits are included below.

9.3 research and rights ledger

Retrieved 2026-10-10. Original instructional synthesis for a simulation-first beginner lesson. Sources are cited and paraphrased; no source screenshots, video, diagrams or long passages are reused. The toy controller, event traces, questions and motion artwork are original. This is not a conformity assessment, legal interpretation or hardware operating procedure.

Instructional synthesis and explicit assumptions

The goal is to evaluate evidence and accessible interaction, not to memorize interface colors. A fictional delivery cart announces station B. Learners trace a cancellation, a terminal result, stale motion feedback, changed scene and deliberate restart. The cancel request blocks new task dispatch locally; it cannot make in-flight physical motion disappear. A fresh zero-speed report plus CANCELED enables a supervisory HELD label in the toy model only. The model's 0.5 s maximum observation age and 0.01 speed tolerance are invented lab constants, not recommended real-world thresholds. A communications-loss state is UNKNOWN and never a claim of a physical stop. Actual applications need application-specific hazard analysis, engineered protection and qualified verification.

The toy policy also requires deliberate confirmation after state/scene checks and uses a new goal ID on resume. This is a teachable policy choice, not a claim that every robot must wait for confirmation before every movement. No human research was conducted; do not claim usability gains, safety certification, universal accessible operation, measured learning gains or virality. Original HH-supplied brand assets and configured ElevenLabs narration are the only non-code production inputs; their existing authorized role is preserved.

Claim-to-assessment mapping

• Q1 goal inference versus path expectation: hri-legibility.
• Q2 non-color and keyboard access: wai-color, wai-keyboard.
• Q3 cancellation acceptance versus completed result: ros-actions.
• Q4 result and measured state separation: ros-actions plus explicitly labeled original model.
• Q5 stale evidence arithmetic: original MAX_AGE policy.
• Q6 restart and old-goal messages: original policy informed by goal identities in ros-actions.
• Q7 software cancel versus emergency protection: ur-emergency, scoped to avoiding equivalence rather than hardware instructions.