Why Household Robots Fail (and How Persistent Object Memory Changes the Game)

Theoretical framing inspired by Stanford AI Index 2026: “Robots still fail at most household tasks, even as they excel in controlled environments.”

The Problem: Controlled Environments Reward Perception; Homes Reward Memory

The Stanford AI Index 2026 highlights a persistent paradox: robots achieve > 90% success on benchmark tasks in lab settings (Pick-and-Place, NAVIGATION in empty rooms), yet drop to < 35% success in real homes [Stanford AI Index 2026]. Why?

Current robotics stacks optimize for frame-level perception accuracy. But household tasks require temporal object cognition: remembering where the keys were placed, tracking the mug through cabinet occlusion, inferring that “the thing under the towel” is likely the remote.

Theoretical framing: This is not a perception problem — it is an embodied memory problem. Human infants develop object permanence around 8 months [Piaget]; robots remain at the “out of sight = out of existence” stage.

FOUR Ways CORE Addresses the Household Gap

1. From Frame-Level Detection → Persistent Object Files (Cognitive Psychology Grounding)

Theory: Kahneman & Treisman’s Object Files theory posits that human vision maintains transient, object-specific representations that integrate features across time. CORE operationalizes this via UUID-level kernels with GRU state.

Implementation example:

Theoretical contribution: This implements the object token hypothesis: identity is maintained not by visual similarity alone, but by a persistent neural representation that integrates new evidence conditionally.

2. From Static Affordances → Dynamic Relational Affordance Inference

Theory: Gibsonian affordances are not static properties of objects, but relations between agent capabilities and environmental structure [Gibson, 1979]. CORE’s GNN enables dynamic affordance inference via relational context.

Implementation example:

Theoretical contribution: Affordances are not pre-labeled; they emerge from the relational graph. A “cup” on a table has high containable confidence; the same cup on a high shelf has high fragile + hard_to_reach (inferred via GNN propagation).

3. From Task-Specific Policies → Generalizable Memory-Conditioned Planning

Theory: Household tasks share latent structure: locate object → navigate → manipulate → verify. CORE’s queryable kernel pool enables a single planning module to condition on memory state rather than relearning per task.

Implemtation example:

Theoretical contribution: Planning is memory-grounded, not perception-grounded. The planner queries “where was X?” before “where is X?”, enabling recovery from occlusion and object displacement.

4. From Episodic Learning → Continual Kernel Adaptation (Lifelong Household Learning)

Theory: Households are non-stationary environments: new objects arrive, layouts change, user preferences evolve. CORE’s kernel pool supports continual adaptation without catastrophic forgetting via uncertainty-gated updates.

Implementation example:

Theoretical contribution: Adaptation is gated by uncertainty: high-confidence observations drive fast updates; ambiguous observations trigger slow meta-learning. This implements the stability-plasticity dilemma [[Grossberg, 1980]] in a principled, deployable form.

To access the full implementation code for the household robots and the COREWorldModel, visit https://coreworldmodel.com