IntSenseIntSense
Sense Humans. Guide Machines.
The human interaction layer that helps machines understand when and how to engage.
A robot that avoids collisions can still frighten, crowd or startle the person in front of it. IntSense decides whether, when and how it should act — on prototypes in R&D and on fleets already on site.
Your fleet already plans routes and avoids people reliably. What nothing decides is whether to approach this person, right now, at this angle — or wait, ask, or back off. That gap is where staff take manual control, where people get startled, and where rollouts stall.
We agree the site, the embodiment and the one interaction that keeps costing you time. Data scope and retention are written down before a single sensor is switched on.
Machines already know how to move. IntSense decides when and how they should engage with the person in front of them — and hands that decision to the stack you already run.
Every time an operator takes control because the robot handled a person badly is a logged event. That count is the first number we move.
Robots that read the room stop freezing in shared aisles and stop making people detour. Cycle times hold closer to the model in your business case.
You leave the pilot with a written benchmark against your current baseline — the evidence your operations and safety leads ask for before a fleet rollout.
The layer we are building: IntSense fuses multimodal human signals into an interaction state, predicts the likely human response to each candidate action, and emits a single decision your controller can accept or ignore. It is a policy layer — never a motion planner.
Voice and intonation, vision, language, distance and optional wearables — captured on device.
Gaze, engagement, intent, crowd density, zone rules and the robot's current task.
P(human outcome | state, context, candidate robot action), with a calibrated confidence.
One high-level call: SLOW · STOP · REROUTE · ASK · ENGAGE · DISENGAGE, with alternatives.
Your planner and controller execute the motion exactly as they do today.
{ "decision": "ASK", "confidence": 0.82, "state": { "engagement": "open", "gaze": "on_robot", "distance_m": 2.4 }, "context": { "zone": "ward_corridor", "crowd": "low", "task": "supply_delivery" }, "alternatives": ["SLOW", "REROUTE"], "expires_ms": 600 }
Every decision carries a confidence and an expiry. Your controller decides whether to act on it, and your existing safety layer always has the final word. Nothing we emit can override a stop.
Each decision is logged with the human outcome that followed it — approach accepted, path abandoned, handover completed. You get a benchmark against your current navigation baseline, not a demo reel.
We sit above your robot brain and below nothing. If your stack can accept a high-level suggestion over a topic or an SDK call, it can run IntSense — no hardware change, no model swap, no vendor lock.
Most teams meet us with a prototype on a bench, not a fleet on a floor — and that is the cheaper moment to decide how the machine will behave around a person. We work with robots still under development: a single unit, a lab, a research programme. No production volume required, no fleet commitment, and the interaction contract is fixed before your hardware is frozen.
A robot that keeps its clearance can still startle someone into a machine aisle, crowd a nurse mid-procedure, or trigger an operator to grab manual control at the wrong second. Those are safety events, and today nothing in the stack decides them. IntSense makes that decision explicit, logged and reviewable — while your certified safety functions keep the final word and nothing we emit can command torque or override a stop.
IntSense emits decisions, not trajectories, velocities or joint commands. Your motion stack and your certified safety functions are untouched, which keeps your existing validation valid.
The full loop — speech emotion, transcription, local language model, response — runs on the compute already on the robot or on a local edge box. No connectivity requirement, no audio leaving the site.
Data scope, retention and use rights are written into the contract before the first sensor is switched on. Site calibration and a privacy and safety audit are part of every deployment.
The policy is the same everywhere; the thresholds, zones and acceptable interruptions are calibrated per site during onboarding.
Mixed aisles where pickers and AMRs negotiate the same lane. The robot learns when to hold back and when to keep moving, so fewer runs end in a manual override.
Corridors and wards where interrupting the wrong person at the wrong moment is the real failure mode. The layer reads readiness before the robot ever speaks.
Approach that reads as helpful rather than intrusive — and a clean disengage the moment the answer is no, without the robot trailing the customer down the aisle.
Embodiments people read socially, where posture, timing and voice carry as much weight as the path. One policy travels across every body you deploy.
Every engagement starts as a paid pilot with a benchmark agreed before the work begins — and you keep that written benchmark whether or not you licence anything. Prices below are list ranges; scope sets the final figure.
Send the embodiment, the site and the interaction that keeps going wrong. We reply with an integration sketch for your stack and a pilot scope — or an honest note that it is too early for your fleet.
Your embodiment, your stack, the interaction that costs you time on site.
Written, specific to your controller — where the decision enters and what it can and cannot touch.
One site, 8–12 weeks, a fixed price and a benchmark defined before we start.
Veronika Kalneus
Founder & CEO, IntSense
Social Intelligence for Physical AI
partnerships@intsense.ai
hello@intsense.ai
A working zero-cloud loop — speech emotion, transcription, local language model and speech — where the inferred human state already changes the interface, the voice and the next action.
Multimodal response prediction and a cross-robot interaction policy, proven on a drone, an AMR and a humanoid across three live sites.
A small team: applied AI and voice analytics on the product side, transformer and speech engineering on the model side, advised on speech AI from inside the industry.