AgentPay / LABS

Human intent.
Robot action.

Exploring how human intent becomes accountable robot action.

Independent experiments in interaction, orchestration and learning from failure.

Concept illustration of a quadruped robot on a terracotta plateau
INTERACTION · ORCHESTRATION · LEARNINGResearch in progress · London

A correction should change the task.

Explore a scripted episode: a person corrects a task, the harness holds execution, checks intent and records the next action.

EPISODE REPLAYSCRIPTED SCENARIO
Schematic facility and task routeAn illustrative robot route between Bay 01 and Bay 02. This is a diagram, not camera footage or a physics simulation. BAY 01BAY 02

SCHEMATIC FACILITY · SAMPLE POSITIONS

Built here

Browser playback, scenario switching, a task state inspector and a downloadable event trace.

What to test next

Interruption handling, correction accuracy and ownership handover on a connected robot.

Evidence boundary

A frontend demonstration with synthetic events. No physical robot, live model or trained policy is involved.

Six authored events · playback positions in seconds

From questions to experiments.

Concept artwork of a quadruped robot observing a greenhouse environment

Embodied fieldwork

Next experiment: a repeatable inspection route, operator corrections and an inspectable episode trace.

Coral and sage ribbons crossing in a sculptural representation of continuous conversation

Continuous interaction

Next experiment: compare turn-taking and duplex interaction on the same correction task.

One research harness.
Different physical environments.

Research concepts · hardware validation remains a next step.

Notice what changed. Make the next step clear.

A quadruped revisits inspection points, flags a visual change and asks a person to confirm the task. The harness records the observation, decision and outcome.

WHAT TO MEASURE Detection quality · intervention rate · time to resolve

A conversation is
only the beginning.

The research question: how should human intent, robot observations and bounded skills meet? This proposed harness separates task decisions from local control, and episode review from policy rollout.

ILLUSTRATIVE ARCHITECTURESELECT A NODE TO EXPLORE ↗
ROBOT FEEDBACK →Versioned rollout ↩

Fast loop: observe → decide → bounded skill. Learning loop: episodes → training → validation.

A missed detection becomes a synchronized episode: video, robot state, action and outcome. Review the cause before changing the policy.

Questions I’m working on.

Rajiv Baskaran.

I build agent software and investigate how human intent becomes reliable, inspectable action.

My current focus is operator interfaces, task orchestration and learning from failure. Physical robot validation is a next step.

Let’s make the next experiment useful.

For conversations about the work, experiments and human–robot interaction.

ALSO BUILT / AGENTPAY

Payment authorization is one part of accountable agent action.

AgentPay docs
AGENTPAY LABS / RESEARCH NOTE