- Chidakashi Robotics
- Chidakashi Safety
Physical AI for Hazardous & Labor-Intensive Work
The robotics foundational models plan and execute actions, gated by Kavach safety before they reach the actuator.
Foundational Models for Physical AI
Chidakashi.
A neuro-symbolic modular platform for physical AI.
Machines that act — and a layer that decides whether they may.
Companion
Mobile
Industrial arm
Quadruped
Underwater
Humanoid
CCTV
Car
Chidakashi: A Modular
Platform for Physical AI
Three classes of foundational model — perception, action, safety — each independently versioned and licensable, each running as a cloud API or inside the control loop on the device.
InputThe world, as sensed
Vision
First-personPerceptionGrounds it
DrishtiVision
ShrutiAudio
SparshTactileActionPlans and generates
KriyaVision-Language-Action policy
Foundation + world modelPrana + KarmaA reasoning core that predicts what happens next and what to do next.SafetyVerifies every action
KavachVerified actuation
ActuatorExecutes it
Industrial arm
Precision cellEvery action alters the environment
Model evolution
Chidakashi Robotics
Kriya
4B → 12B
A unified family of Vision–Language–Action policies for contact-rich manipulation.
Karma
4B → 16B
World model — predicts how objects, people and other agents will respond before the machine acts.
Prana
9B → 37B
Flagship omni-conditioned foundation model with a dual-brain architecture — one family that perceives the world, anticipates what comes next and acts across any embodiment.
Generation 1
Kriya and Karma, live today — single-arm cells, verified at control rate
Uses vision, proprioception and wrist force–torque to distinguish visually similar but physically different conditions, such as successful insertion versus jamming, stable contact versus collision, or a secure grasp versus an unexpected change in load.
Generation 2
Prana world action model arrives
Adds multi-arm coordination, richer scene-conditioned manipulation and distributed touch from fingertips, finger links and palms. Tactile readings are attached to nodes in the hand’s kinematic graph and represented as embodiment-independent contact tokens containing location, pressure, normal force, shear and deformation. This allows contact experience to transfer across grippers and dexterous hands with different finger counts, joint structures and tactile layouts.
Chidakashi Kavach: The World’s
Safest Brain for Physical AI
Independent Safety Benchmarks — 1.00 prompt safety and 0.91 response safety, on Kavach v2 models already in production.
More accurate
Beats every flagship model in the set on both measures.
Inside the loop
Verdicts at control-loop latency, not datacenter latency.
Harder to break
The margin widens under multi-turn escalation and injection.
Prompt safety
Score0.800.901.00
Chidakashi Kavach1.00
Claude Haiku 3.50.96
GPT-4.0 Mini0.96
Grok 3 Mini0.94
Gemini 1.5 Flash0.93
Response safety
Score0.800.901.00
Chidakashi Kavach0.91
Gemini 1.5 Flash0.86
Claude Haiku 3.50.85
Grok 3 Mini0.84
GPT-4.0 Mini0.83
Kavachv3
Same trend, larger opponents
Same benchmarking trend against large frontier reasoning models — and it holds in multimodal domains, on visual and audio guardrails, not text alone.
Kavachv4
Into physical embodiment
Extends into physical embodiment — the same verification semantics over embodied action, every morphology.
How Every Chidakashi Model Improves
The Self-Learning Flywheel — Reinforcement Learning in the Loop
Every improvement is earned by reinforcement learning against audited outcomes of the fleet’s own experience. One learning loop, every model class.
1
Real-time · Edge node
Deploy
Perception, action and safety models execute inside the control loop and emit a signed trace of every inference.
2
Asynchronous · Cloud
Audit
A high-reasoning world model replays that trace on cloud infrastructure, adjudicating what the edge model perceived, planned and permitted.
3
Simulation · Lab
Improve
Confirmed misses are reproduced in simulation in generated worlds; synthetic rollouts train candidate models through reinforcement learning and self-play.
4
Autonomous · Test bench
Validate
Every candidate is run against an autonomous sim-to-real test bench — the same scenario executed in simulation and then on physical rigs, with no operator in the loop.
5
Release-gated
Release
Improvements ship only as versioned, signed artifacts, and only after regression against the full audited trail plus adversarial testing and red-teaming.
Chidakashi Data Lab
Multiple Concurrent Streams, Each Closing a Different Data Gap.
Stream 01
Chidakashi Data Factory
40+ end tools, 100+ tasks and subtasks captured with ego-centric cameras and tactile instrumentation. Every run is scored.
Stream 02
Manufacturing & Logistics
50+ instrumented stations streaming time-synchronised video, depth and force telemetry off work already happening.
Stream 03
Home & Consumer
Capture across homes for the household distribution, including tactile-sensing glove and 3-camera egocentric rigs.
Stream 04
Digital Twin & World-Model Synthesis
Mirrors real premises and multiplies every trajectory into simulated variations, de-risking sim-to-real.
Chidakashi in Enterprise and Industry Deployment
- Chidakashi Audio
- Chidakashi Vision
- Chidakashi Safety
Large Scale Education Deployment
Built on Shruti’s speech understanding. The system scores fluency, tracks per-child progress across the school year and surfaces classroom-level insight to teachers — with every interaction passing through Kavach realtime multi-turn trajectory analysis.
- Chidakashi Audio
- Chidakashi Safety
Kavach Layer in a Global Gaming Major
Running inside the customer’s own stack, Shruti and Kavach powering NPCs — COPPA and safety compliance carried by the Kavach safety platform.
- Chidakashi Audio
- Chidakashi Vision
- Chidakashi Safety
A Decade in Half a Million Homes
The proving ground: ten years shipping to children.
500K robots deployed across 140+ countries, and the #1 kids robot brand in the US. Multimodal Kavach safety across conversation and generated-image evaluation.
Kavach v3: Vision-Led Multimodal Safety
LikenessNo consent. Blocked.Content & generated media safety
36 content and generated-media policies built in
Harmful and sensitive imagery. Generated media is not just deepfakes but is verified safe, faithful, physically coherent and fit for use.
EscalationVoice and posture agree.
ContextLoud. No weapon, no contact.Multimodal AI safety
Severity, not keywords
One foundational model jointly analyses speech, paralinguistics and the accompanying image or video.
ProximityWorker inside the forklift envelope.Outside-in camera
Facility-Wide Sensing and Safety
An outside-in safety agent expands awareness by communicating with sensors and cameras placed throughout the facility. 60+ prebuilt alerts.
Kavach v4: Physical AI Safety
01 · Manipulators · cobots
Manipulation Safety
Pre-execution checking on force, velocity, reachability and workspace occupancy; power or force-limiting and protective stops on contact.
02 · Humanoids · quadrupeds
Locomotion & Whole-Body Safety
Whole-body motion, balance and force bounded around people; safe-stop and kill-switches; envelopes that hold under sensor noise, occlusion and unexpected contact.
03 · AMRs · drones
Navigation & Flight Safety
Collision-free navigation in dynamic shared spaces; bounded speed and proximity; safe-stop and fallback on sensor loss.
04 · Outside-in
Outside-In Safety
Kavach v4 runs a two-brain system: outside-in for early warning, onboard for the metric bound that gates the actuator. Disagreement is never averaged — it is scored as uncertainty, and the envelope tightens.