Robotics Is a Constraint-Satisfaction Problem

The dominant intuition in machine learning is that capability is a scaling problem. Bigger model, more data, more compute — and reliability follows. For robots operating in the physical world, this intuition is wrong, and the way it is wrong is instructive. Robot intelligence is better understood not as a function to be approximated but as a set of constraints to be satisfied simultaneously.

Robots don't degrade gracefully

A language model that is slightly undertrained gives slightly worse answers. Performance degrades smoothly. A robot is different. When a critical condition is missing, the system does not get a little worse — it fails stochastically. The arm drops the object, repeats a failed grasp, or enters a trajectory it cannot recover from. There is no partial credit.

This is the signature of a constraint-satisfaction problem, not an optimization problem. A constraint-satisfaction problem (CSP) is classically defined as a triple of variables, their domains, and a set of constraints that must all hold. A solution is an assignment that satisfies every constraint at once. Violate one and the entire assignment is invalid. That “all at once” structure is exactly why robots collapse rather than coast: feasibility is not an average, it is a conjunction.

The five constraints

A useful decomposition that has been gaining traction in the embodied-AI community holds that a robot is stable in the real world only if five conditions are satisfied together:

• Memory — degradation prevention. The policy must not keep re-entering failed trajectories. Without a mechanism that treats no-repeated-failure as a hard constraint, errors propagate and compound.
• Consistency — physical causality. The model must represent cause and effect under physical law, not merely fit patterns in pixels. This is the only real entry point into the physical world.
• Embodiment — the structure of the action space. If the policy's action representation doesn't match the robot's degrees of freedom and dynamics, the achievable intelligence is bounded from above no matter how much data you add.
• Data — distribution coverage. No single source covers the combinatorial explosion of real-world situations. Coverage gaps show up as failures outside the training distribution.
• Planning — temporal abstraction. Long-horizon tasks require reasoning over subgoals, not myopic next-action prediction.

The important claim is not the list — it is that these are not independent. They form a loop: consistency shapes representation, representation defines what data is usable, data bounds the policy, the policy determines interaction, and interaction feeds back to close the data loop. Satisfy one and the others become easier. Miss one and the loop breaks.

Why scaling alone fails

Modern generative models are unconstrained function approximators. They minimize an average loss over a corpus, with no hard feasibility condition enforced at inference. The output can be anything plausible in the training distribution — including physically impossible actions and confidently repeated failures. Pouring in more data does not fix this; it averages conflicting strategies into a blurrier policy. That is why “larger model” does not reliably mean “more reliable robot,” and why “more data” does not reliably mean “better generalization.”

Constraint satisfaction inverts the objective. It treats feasibility as primary and turns learning from pure function-fitting into something closer to pruning the space of trajectories — removing the ones that violate physical, memory, or embodiment constraints. The field has fifty years of machinery for exactly this: constraint propagation, motion planners such as RRT* and CHOMP, symbolic task planning with PDDL and hierarchical task networks, inverse-dynamics and whole-body control, and modern hybrid solvers. The frontier question is how to fold those hard constraints into learned policies rather than bolting them on afterward.

What this means for data

If reliability is constraint satisfaction, then the value of a dataset is not the number of hours it contains — it is how well it lets a model satisfy the constraints. Two implications follow directly.

First, consistency has to be in the data. A model can only learn physical causality from samples where cause and effect are correctly corresponded in time. Temporally misaligned multi-modal data actively teaches the wrong causality — which is why the synchronization layer is not a nicety but a precondition for the consistency constraint. Second, coverage and separability both matter. Data has to span enough of the real distribution to bound the policy, while staying labeled and structured enough that the model can tell situations apart rather than averaging them.

This is the lens we build under. Validated, integrity-checked, contact-rich episodes are not “more footage.” They are the substrate that lets a model satisfy consistency, embodiment, and coverage at the same time. Failure, in the end, is not noise. It is a violated constraint — and the only durable fix is to build the data and the system so the constraints can hold.