01Foundation models for physical AI
Machines that act — and a layer that decides whether they may.
Neuro-symbolic perception, action and safety models. By Miko.
Companion
Mobile
Industrial arm
Quadruped
Underwater
Humanoid
CCTV
Car
0
Robots in real homes
0
Hours moderated
0
Countries
0
Patents · 10 granted
02The platform
Perception, action and safety — versioned and licensable on their own.
Vision
First-person
Industrial arm
Precision cellPerception
Vision, audio and touch become latents a rule can be written against.
Action
The World-Action Model predicts the next frames and the next action together.
Safety
No prompt, persona or training can address it. Zero learned components on the decision path.
Because Kavach is never part of the generation context, there is nothing in a conversation — or a task — that can argue it into a different verdict. Policy compiles to a formal constraint specification and is decided clause by clause, so every verdict comes with the clause that produced it and a signed, reproducible record.
03The proving ground
Children. 100+ engineers building for real-time edge nodes the whole time.
0
Robots deployed
0
Cumulative revenue
#1
Kids robot brand
Retail & licensing · eight global brands
04The gap
Every layer of physical AI has incumbents fighting over it — except the one that decides whether a machine may act.
Contested · and others
Contested · and others
Contested · and others
Contested · and others
Contested · and others
Contested · and others
Contested · and others
Contested · and others
Contested · and others
Safety & verification
No incumbent
Where the field actually is
Can walk, lift, navigate
SolvedCan answer questions
SolvedAllowed into a home
UnsolvedAccepted in a hospital
UnsolvedLeft alone with a child
Unsolved“A robot that can carry boxes but makes humans instinctively uncomfortable is not a product. It is a liability.”
A humanoid kicked a child at an amusement park — China, June 2024.
The documented incidents05Benchmarked & published
1.00prompt safety and0.91response safety, on models already in production.
Prompt safety
Response safety
Beats every flagship model in the set on both measures.
Verdicts at control-loop latency, not datacenter latency.
The margin widens under multi-turn escalation and injection.
Benchmarked and published
The scores above.
Vision, audio, larger opponents
Holds against frontier reasoning models.
Into physical embodiment
The same semantics over embodied action.
06In the field
Corporate partnerships, government mandates, real environments.
Robots on the floor, monitored for quality and safety.
The Chidakashi action stack plans and executes across robots on the line, Chidakashi Perception drives continuous industrial monitoring, and every action is gated by Kavach before it reaches an actuator.
Assessed daily. Every interaction moderated by Kavach.
Built on Shruti’s speech understanding. Each child builds an egocentric understanding model that maps progress across the school year and surfaces classroom-level insight to teachers.
Shruti and Kavach powering NPCs, approved portfolio-wide.
Running inside the customer’s own stack. COPPA and child-safety compliance carried by the models themselves rather than a downstream filter — a regulatory and brand priority for the customer.
07Chidakashi Data Lab
Most of it arrives as a byproduct of operations already running.
Stream 01
40+ tools, 20+ household tasks, scored every run.
Allo-centric cameras and tactile instrumentation across physical-AI embodiments: human egocentric video, force-torque, teleoperation and autonomous rollouts. Scoring drives success-rate optimisation, autonomy-level assessment and safety testing.
Stream 02
50+ instrumented stations. Real technique, real clutter.
Time-synchronised egocentric video, depth and force telemetry off work already happening — the long-tail distribution industrial policies have to generalise over, segmented and labelled to the process step.
Stream 03
The household distribution. Gloves and 3-camera rigs.
The hardest distribution to buy and easiest to get wrong: variable lighting, clutter, soft objects and people moving unpredictably through the workspace.
Stream 04
Every real trajectory, multiplied in simulation.
Mirrors real premises to de-risk sim-to-real. Video-generation and world models synthesise additional egocentric viewpoints aligned to real captured data.
08The loop
The fleet’s own experience trains the next version. One loop, every model class.
Deploy
Edge, in the loop
Audit
Cloud replays the trace
Improve
Retrain in simulation
Validate
Sim-to-real bench
Release
Signed, versioned
One loop ·all eight modelsRelease-gated throughout
Every robot will need a reason to be trusted.