World models · Autonomous agents · Planning
Plans longer.
Learns from less.
Continuum is a compact world model for autonomous agent training. It maintains a probabilistic state of the environment, predicts how that state changes under each action, and selects actions based on those predictions — reaching 50-step horizons from a fraction of the real interactions.
The problem
What limits training
autonomous agents
Cost of real interaction
Experience is collected on a physical device — bounded by time, equipment wear, and the irreversibility of some actions.
Tracking state over time
Over long action sequences the agent must remember what it already changed: opened doors, spent resources, moved objects.
Rare scenarios
The situations where reliability matters most appear in data least often. Collecting enough of them is expensive.
Partial observability
Decisions are made from partial information — the rest is hidden by walls, corners, and sensor limits.
The approach
A model that thinks ahead
Instead of predicting future frames, Continuum maintains a compact symbolic state — objects, properties, and relations. The transition is conditioned on the action: the model answers "what changes if I do A", not "how does the image continue."
Hidden parts of the environment stay as distributions. What is not observed is kept as a probabilistic hypothesis rather than discarded. Uncertainty enters action selection — an information-gathering step has value when it lowers risk ahead of an expensive move.
Step horizon targeted. Reactive agents typically fail past 25.
Fewer real episodes to target quality (hypothesis; to be confirmed on benchmark).
Success rate target at 50 steps vs ~40% for the no-state configuration.
Architecture
Observation encoder
Converts raw observation into a compact state — objects and relations, not a pixel grid.
Action-conditioned transition
Predicts how state changes under a specific action, with uncertainty tracked through the transition.
Uncertainty estimate
Hidden cells stay as distributions. Unknown regions are hypotheses, not missing data.
Planner — MPC / beam search
Simulates candidate sequences in latent space, selects based on predicted outcome and uncertainty cost.
One interaction → 10²–10³ simulated variants
Each real step feeds the dynamics model. The policy updates from the generated rollouts before the next real interaction.
Prior art
What the literature shows
World-model surveys published in 2026 name three open problems: error accumulation over long rollouts, sim-to-real transfer, and the absence of a shared evaluation protocol. Continuum works on the first and the third.
DayDreamer · 2022 · arXiv:2206.14176
Real time to learn walking
A world-model agent trained a real quadruped to walk in approximately one hour of physical interaction — without any simulation.
WorldSample · 2026 · arXiv:2607.02431
Training steps in manipulation
59% fewer training steps and 28% more successful episodes over model-free baselines on robotic manipulation tasks.
Both results come from specific environments and specific teams. For Continuum they set an order-of-magnitude reference that still has to be verified on our own benchmark.
18-month plan
Five stages,
five verifiable artefacts
Every stage closes with code, results, and a report an outsider can check. No stage extends on a negative result — the refutation criterion is fixed before experiments start.
M 0–3
Environment & baselines
BeliefWorld API, task set, reproducible configurations A–C, metrics fixed.
M 4–6
Dynamics
Transition model, action-conditioned prediction, latent vs pixel comparison, ablations.
M 7–9
Belief state
Probabilistic state, calibration measurement, technical report + recorded demo.
M 10–12
Horizon & transfer
Hierarchical planning, new map testing, sample-efficiency measurement.
M 13–18
External environments
Industrial simulator integration, partner runs, SDK + API, next-round prep.