IRONHEART.AI
PHYSICAL AI FOR A LIVING WORLD
A body can move.
A mind can relate.
We’re building the cognitive runtime that brings perception, voice, memory and personality into one continuous interaction.
Follow a single exchange through the layers of a cognitive system. An interactive explanation of the architecture—not live robot data.
01 / PERCEPTION & ATTENTION
A useful interaction starts before the first word. A camera, a microphone and the surrounding context help a robot decide what deserves its attention.
02 / VOICE & TURN-TAKING
Listen while speaking. Yield when interrupted. Let a pause mean something. Timing is part of behaviour, not an afterthought.
03 / MEMORY, KNOWLEDGE & IDENTITY
What was just said is not the same as what should be remembered. Give each runtime its own knowledge and deliberate rules for identity, consent and retention.
04 / PERSONALITY & MODEL ROUTING
Warm or matter-of-fact. Playful or precise. Compose the voice, language and boundaries that fit this robot’s role—then inspect the result.
05 / EMBODIMENT & OBSERVABILITY
A reply can guide attention. An action needs permission. Keep the conversation, the hardware state and the operator’s view connected—without confusing intent with execution.
Perception → Voice → Context → Response → Approved action
Explore RGB and infrared views, face-presence signals and gaze targets. In a connected system these signals require a calibrated sensor adapter. The Studio preview uses synthetic scenes, not a live camera.
The control room separates voice activity detection, streaming speech recognition, model prefill, response chunks, speech synthesis and playback. Inspect a turn to see how interruption cancels a sample response. Timings are illustrative, not performance claims.
Separate session context from persistent memory. Scope a record to a workspace, robot and end-user identity; make consent, retention and deletion explicit. The account-gated demonstration stores sample data in this browser and must not be used for confidential information.
Try the personality mixer, inspect the compiled prompt and compare a draft with a published demo version. Model routing and fallback are configuration concepts; a real deployment needs validated models and service contracts.
Inspect the boot pipeline, service health, gaze and motor readback. A motor fault stays a fault after acknowledgement. Recalibration and restart buttons in Studio are dry-run simulations; no command reaches a physical robot.
Our work with Captcha has taken us into classrooms and public events. These are specific field demonstrations—not claims of universal autonomy or production-scale deployment.
Captcha took part in simulated oral Abitur examinations at Willms-Gymnasium. NDR's report, broadcast on 17 April 2026, follows a mock examination with student Miriam Hollmann.
The exercise brings the cognitive-system problem into focus: follow an answer, maintain the examiner's role and keep the exchange moving. It does not establish that robots can replace teachers or deliver accredited assessments.
Watch the NDR report ↗Captcha has appeared at CES in Las Vegas, meeting and speaking with visitors. The team-provided film documents the robot in a public environment.
For an interaction system, a visitor's silence matters too. Recognising non-engagement, choosing when to speak and knowing when to stop are behaviours to evaluate—not assumptions we can prove from a short video.
Explore the research field record ↗Design an interaction system for a particular body, environment and purpose. A classroom tutor should not behave like an exhibition host.
Explore a reusable cognitive layer while keeping hardware-specific sensing and control explicit.
Define knowledge, behaviour and data boundaries for a use case, then validate them with your engineering team.
Explore the actual Studio demo before creating an account. Three views of how a cognitive runtime takes shape.
Product preview · Sample data · No robot connection
Adjust warmth, humour, curiosity and formality. Save a draft, inspect the compiled prompt and publish a demo version. Try the effect on scripted replies in Playground.
Configuration is stored locally in your browser. No model is running.
Add approved sample notes or text files, then ask a question in Playground and inspect the matched source. Explore memory and identity policies as separate modules.
Demo retrieval uses local keyword matching, not a production embedding service. Use sample text only.
Replay a sample conversation and inspect speech detection, transcription, generation and speech synthesis on one timeline. The control room also includes perception, service status and a searchable journal.
Transcripts and timings are simulated. This is not live telemetry or a performance benchmark.
Interactive access requires registration. Self-service signup is being prepared.
Request demo access from the team, or sign in to your existing workspace. No payment required.
Studio is a demonstration, not a checkout. Cloud API access and hardware deployments are discussed directly with the team. The figures below describe the proposed monthly Cloud pricing.
Indicative plan for scoping a future API evaluation.
Indicative plan for scoping a future API evaluation.
Indicative plan for scoping a future API evaluation.
Indicative pricing, subject to a written offer. Credit conversion, non-voice usage, taxes, module availability and service commitments must be confirmed for your integration. No charges are made by the demo.
Discuss endpoints, runtime capabilities and usage limits with the team. Production keys are not issued by the demonstration.
Define target hardware, languages, selected modules and disconnected operation. Installation and licensing are arranged separately.
Scope infrastructure, data handling, security responsibilities, updates and support through an enterprise discussion.
Our thesis: the value of a social robot depends on how coherently its capabilities work together. We are turning hands-on research into a modular product experience.
Captcha demonstrations in education and public settings, with linked reporting and a documented research history.
A Studio demonstration for composing behaviour, inspecting configuration and exploring a versioned runtime workflow.
Validate connected modules, hardware-specific deployment and commercial service delivery with selected partners.
For robot OEMs, system integrators, research teams and investors. Tell us about your hardware, use case and the interaction you want to make possible.
IronHeart is developed by Hidoba Research. Program membership does not imply investment, certification or product endorsement.
