01Robotics foundational models
Chidakashi Robotics.
Three models, one action stack — and every action verified byKavachbefore it reaches an actuator.
Policy · VLA
Single- and multi-arm control at the edge, 30–120 Hz.
World-action model
Obstacle awareness and correction inside the loop.
Omni-conditioned
Dense perception paired with a Karma core.
02Architectural research
Four deliberate departures.
Differences in how the models are trained and represented — not in how much data is thrown at them.
01 Handoff is the forcing function
Coordination learned inside one set of weights, not negotiated at runtime.
The field vs Chidakashi
The field:Schedulers negotiate at runtime and break at handoff: one manipulator occludes the object from the other’s camera at the instant transfer must succeed. A scheduler passes a coarseready; it cannot pass contact state, grip confidence or the tactile read of release.
Chidakashi:Each agent is conditioned on the other’s predicted future state, so the handoff is learned inside a single set of weights. The transferring agent’s visual and tactile signal is a first-class conditioning input, not a delegation tier.
02 One trunk, two heads
A new body adapts on kinematic structure, not on robot identity.
The field vs Chidakashi
The field:One action representation at a fixed dimensionality, so every new morphology needs a new head.
Chidakashi:A shared trunk with two heads selected at inference — an embodiment-agnostic task-space head, and a superset joint vector masked per platform. Because both read one trunk, egocentric human video improves representations under joint-level control it can never supervise.
03 Trained on hardware that drifts
Non-ideal actuation sits inside the training distribution.
The field vs Chidakashi
The field:Policies assume idealised, well-calibrated hardware — and no public benchmark measures degradation under actuation drift or sensor loss.
Chidakashi:Non-uniform motion, latency skew and drift with wear are in the distribution, so the policy learns closed-loop compensation rather than assuming a nominal plant. Under partial failure it detects the loss from state divergence and hands the task off safely.
04 A fleet that remembers
One robot’s failed grasp becomes every robot’s prior.
The field vs Chidakashi
The field:Policies are memoryless within a task — every out-of-distribution moment is re-derived from scratch.
Chidakashi:Two memories learned together: episodic (what happened here?) and motor/value (which actions worked?). Both are federated across the fleet, so failures concentrate experience exactly where near-contact blindness lives.
03Roadmap
One stack, four generations.
Kriya, Karma and Prana advance together. Generation 1 is live today.
Generation 1
Single-arm cells, verified at control rate
Kriya drives the edge; Karma corrects inside the loop.
What changes
Action requests decode continuous chunks at 30–120 Hz, every chunk verified before execution, while Karma supplies immediate obstacle awareness and short-horizon motion prediction.
Generation 2
Prana arrives; two arms learn handoff
Joint embodiment becomes a training modality.
What changes
Prana pairs omni-perception with a Karma core. Kriya adds multi-arm coordination and closed-loop re-planning of contact-rich pipelines; Karma tracks object state and anticipates scene change. Each agent is conditioned on the other’s predicted future state, so the handoff is learned inside the model.
Generation 3
Long-horizon work, more than one body
The dual action representation lands.
What changes
Kriya decomposes long-horizon work over subgoal latents with learned recovery; Karma reasons causally over longer horizons and predicts multi-agent behaviour; Prana pairs dense omni-perception with a 16B Karma core for multi-embodiment control. Trained with non-ideal actuation in the distribution.
Generation 4
One stack across bodies — and a fleet that remembers
Federated episodic-motor memory closes the loop.
What changes
Kriya transfers policy across embodiments with dexterous multi-stage manipulation; Karma reasons at simulation depth over rare events; Prana runs dense omni-perception on the flagship Karma core and distils back into lightweight action heads. Capability compounds across unitsshipped, not units trained.
04Spatial memory
Where becomes an index
into what happened there.
Objects, events and learnt bounds written into one persistent map frame.
“Bring the tray” is not a perception problem when the object is out of view. It is retrieval.
What the map carries
- Objects as persistent memory— short-term entries decay with observation; long-term entries persist and are corrected on contradiction, so a moved object updates rather than duplicates.
- Behaviour and time— not only what is where, but what happens where. A cell is traversableandhistorically contested at shift change.
- Retrieval— image geolocation returns where a scene was seen; run in reverse, it returns what episodes were recorded here and what has worked before. This transfers across machines, so a room one robot has never entered may be a place the fleet already knows.
- Learnt safety bounds— geofences are commissioned by hand today. Here they are revised continuously from observed traffic, near-misses and human occupancy. Localisation confidence is itself a safety signal: drift means slow, stop, re-localise.
05Active research
Four programmes, already running.
Not a roadmap. Models in training, measured against live deployment data.
Training in imagination
Policies train against imagined consequences, not replayed logs.
Detail
Karma generates action-conditioned rollouts — future frames and actions decoded in one forward process. Rare events are sampled deliberately instead of waited for: slip, occlusion, near-collision, recovery. Every rollout is physics-checked before it reaches the trainer.
Web-scale egocentric learning
Contact events recovered from web video. No annotation.
Detail
Hand pose, contact events and task boundaries are recovered automatically — no wrist markers. Recovered action units carry embodiment-agnostic behaviour into the shared trunk, so pretraining scales past what any fleet can record.
Minimal teleoperation
A handful of demos calibrate. The policy generalises the rest.
Detail
Teleoperation is the most expensive data robotics has, so it isn’t spent on volume. Every operator intervention is logged as a preference signal — a labelled failure, not merely a rescue.
Emergent behaviour
Zero-shot transfer, treated as a measurement not an anecdote.
Detail
Zero-shot instruction following, reactive error recovery and cross-embodiment transfer appear without task-specific training, because all three are aligned in one trunk. Held-out benchmarks over unseen objects, unseen goal geometries and perturbed scenes — a capability counts only when it survives out of distribution.