Introduction

General-purpose robotic intelligence depends on data coverage and distributional alignment, not model scale alone. When systems encounter scenarios that are out of distribution, generalization is limited and performance degrades. Robust generalization therefore requires training on interaction data that progressively covers the tasks, environments, and physical constraints encountered in deployment, achieved through a structured progression across adjacent tasks. This motivates a deliberate approach to how real-world data is collected and expanded over time.

Generalization Degrades Under Large Distributional Jumps

Robotic systems generalize reliably within domains that are well represented in their training data, but performance degrades sharply under large distributional jumps—for example, when transferring directly from industrial manipulation to humanoid household tasks. In such cases, policies encounter out-of-distribution interaction regimes involving unfamiliar contact dynamics, temporal structure, and environmental uncertainty, leading to miscalibrated perception-action mappings and missing recovery behaviors. Robust generalization therefore depends on maintaining continuity between training and deployment distributions, rather than attempting direct transfer across structurally different environments.

A Strategic Path to Generalization

We achieve this continuity by deliberately structuring how real-world interaction data is collected over time. Instead of maximizing task diversity upfront, we expand the training distribution gradually through adjacent tasks that reuse existing interaction primitives while introducing controlled increases in complexity. Each stage extends the model's experience along specific axes—such as coordination, uncertainty, or temporal horizon—while preserving distributional overlap with prior data. This strategic expansion allows the system to incrementally acquire transferable knowledge of real-world physics, system dynamics, and interaction structure, enabling stable and compounding generalization without abrupt performance degradation.

Stage 1: Warehouse pick-and-place as a controlled anchor domain

Warehouse pick-and-place provides an ideal starting point for generalization because it is interaction-dense, repeatable, and safety-bounded, yet already spans a meaningful portion of the manipulation manifold. The task exposes robots to diverse object geometries, packaging materials, clutter patterns, and grasp affordances, while operating under strict throughput and reliability constraints.

Crucially, this domain enables large-scale real-world data collection with high action fidelity, allowing models to learn robust visuomotor alignment, contact dynamics, and recovery behaviors. From a data perspective, it forms a dense core in the action embedding space, onto which adjacent tasks can be incrementally attached.

This domain offers:

• High-volume, repeatable real-world interaction data
• Clear success metrics and dense supervision
• Strong coupling between perception and manipulation

From a learning perspective, warehouse pick-and-place densely samples:

• Object-centric perception under clutter
• Grasping and placement primitives
• Closed-loop error recovery
• Long-tail object distributions

These primitives are directly reused in returns handling; without robust grasp recovery, pose correction, and clutter handling learned here, more complex multi-object or bimanual tasks would fail catastrophically.

Stage 2: Adjacent expansion — returns handling and dual-arm coordination

Returns processing naturally extends warehouse picking while introducing new interaction regimes rather than entirely new primitives. Tasks such as unpacking parcels, sorting heterogeneous items, and performing coarse quality checks require increased perception under uncertainty, deformable object handling, and error-tolerant reasoning.

Dual-arm setups add a critical new dimension: bimanual coordination and role specialization, expanding the action space without abandoning previously learned affordances. Importantly, the underlying data distributions—objects, packaging, logistics environments—remain partially overlapping with warehouse operations, enabling effective transfer while forcing the model to generalize beyond single-arm, single-action sequences.

• Deformable packaging and unknown initial states
• Multi-object reasoning within a single episode
• Bimanual coordination and handover dynamics

Bimanual coordination and object-to-object relationships learned here are prerequisites for kitting and sequencing, where correctness depends on relative placement, ordering, and coordination rather than isolated actions.

Stage 3: Kitting and sequencing in automotive logistics

Kitting and sequencing tasks introduce temporal structure and goal-conditioned manipulation, requiring the robot to reason over ordered sets, intermediate states, and tight tolerances. While still grounded in industrial environments, these tasks shift the learning signal from isolated grasps toward long-horizon consistency, precision placement, and constraint satisfaction.

From a generalization standpoint, this stage expands the dataset along the planning and memory axes, rather than merely adding new objects. The resulting data bridges the gap between reactive manipulation and task-level execution, a prerequisite for more complex forms of embodied intelligence.

• Order-sensitive task execution
• Tighter tolerances and placement precision
• Stronger coupling between perception, planning, and action

Long-horizon consistency and sequencing are necessary for manufacturing, where actions must respect process dependencies and partial assemblies.

Stage 4: Manufacturing and In-Process Interaction

Manufacturing environments extend prior domains by introducing process coupling and contact-rich manipulation. Robots must interact with partially completed assemblies, tools, and fixtures, often under tight force, alignment, and accuracy constraints. Actions are no longer isolated but embedded in a process context, making the action distribution more structured and failures less forgiving. This demands higher-quality representations of state, intent, contact dynamics, and uncertainty.

This stage enriches the data corpus with causal, contact-driven interaction data, enabling models to move beyond pattern matching toward understanding how sequences of actions transform partially assembled systems in predictable ways.

• Contact-rich assembly and insertion tasks
• Tool usage and fixture-based interactions
• Process-dependent tolerances and constraints
• Semi-structured but evolving environments

Learning process constraints and tool-mediated, contact-rich manipulation is a prerequisite for operating in semi-structured public environments, where errors carry higher cost and successful behavior depends on anticipatory, rather than purely reactive, control.

Stage 5: Store replenishment and semi-structured public spaces

Store replenishment tasks push robots into semi-structured environments with higher visual variability, dynamic obstacles, and weaker environmental priors. While still commercially constrained, these settings introduce human presence, occlusions, and frequent distribution shift.

At this point, the model's success depends not on task-specific tuning, but on its ability to adapt policies online, leveraging the accumulated diversity of prior interaction data.

Including:

• Open-world layouts
• Frequent distribution shifts (seasonality, planograms)
• Safety-critical interaction with humans

Exposure to human-facing, partially predictable environments prepares the system for household and healthcare domains, where structure is minimal and safety requirements dominate.

Stage 6: Household and healthcare domains

Household and healthcare settings represent the culmination of this trajectory: unstructured environments, high object diversity, deformables, safety-critical interactions, and long-tail edge cases. Importantly, these domains are not tackled directly, but approached with a model pretrained and post-trained on a rich lattice of adjacent tasks.

Because the system has already explored large portions of the manipulation, perception, and coordination manifolds, these domains become a question of data density and alignment, not architectural reinvention. These domains include:

• Highly diverse objects and environments
• Low repeatability and sparse supervision
• Strong safety, reliability, and trust requirements