We build causal world models for physical simulation and robot learning — agents that reason about what an action will cause before taking it, instead of imitating demonstrations. The engine is live at app.decivine.com.
one causal engine,
simulation and control
fewer training samples
to reach target skill
classic control
benchmarks solved
The platform builds a causal world model from your data, then predicts, answers counterfactuals, and lets a population of causal agents tune the model automatically. This is the engine we are now pointing at simulation and robot data.
Structured data or documents in — the platform extracts causal structure and builds a queryable graph. Intervene on any node and read off the downstream consequence.
Ask “what happens next?” or “what would happen if X changed?” — answered in one forward pass with the full causal chain in plain sight, restricted to verified causal edges rather than correlation.
A population of causal agents trains on the world model, probes it where it is least certain, and writes what it learns back in — accuracy compounds with every pass.
Once your graph exists, the platform extends it automatically — pulling in related entities, relationships and supporting evidence, so the model grows without manual curation.
We walk design partners through the live platform on their own simulation or robot data — building the causal graph and running interventions in the session, not a scripted demo.
Book a live walkthrough →Prefer to explore first? The platform is open at app.decivine.com.
Two markets, one engine. Robot learning is the wedge — every action is an intervention and every real trial is slow and costly, which is exactly where causal beats imitation. Physical simulation is the same buyer, out of the same budget line.
Agents that imagine an action’s outcome before committing — planning with counterfactual rollouts inside the world model, and reaching the same skill from a fraction of the training data.
First design partnersEvery action is an intervention. Causal agents discover the structure of their environment and reason over it — better credit assignment, less confounding, far greater sample efficiency.
ValidatedA causal world model of a physical system — contacts, friction, fluids, structural mechanics. Replace an expensive solver sweep with a single forward-pass query: “what if this parameter changed?”
On the roadmapWe would rather be precise than impressive. Here is exactly what exists today, and exactly what it does not yet cover.
Causal world model, prediction, counterfactual QA and auto-tuning causal agents — running now at app.decivine.com
Classic control benchmarks — CartPole, Pendulum, MountainCar, LunarLander — matched the baseline policy at ~2–3× fewer training samples
Counterfactual queries answered in one forward pass — no chain-of-thought, fast enough to sit inside a live control loop
Those four are simulated benchmarks, not physical hardware, and the result is a research finding separate from the live platform. Next milestone: contact-rich manipulation on real hardware with a design partner.
Early stage. Core validated, platform live. We’re signing the first design partners in robot learning and physical simulation — bring your data and we’ll run a real counterfactual, not a demo.
Causal world models and the agentic systems that enrich them — one engine for physical simulation and robot learning.