Robotics has reached a point where isolated capability is no longer the limiting factor. Robots can grasp, walk, open doors, and follow short instructions with growing reliability. What continues to break down is continuity. The moment a task stretches across rooms, objects, and time, autonomy fractures. Planning resets. Context is lost. The system stops being a system.
The table-to-dishwasher task marks a different threshold. Not because it looks impressive, but because it holds together.
For Alper Canberk, the central challenge of home robotics is not mechanical elegance or model size, but continuity. As the founding Director of Research, Robot Learning & Foundation Models at Sunday Robotics, whose recent public launch out of stealth reshaped how the industry thinks about robotics data collection, Canberk works at the intersection of embodied AI, large-scale generative modeling, and real-world deployment. In this role, he helps define how autonomous systems move beyond short demonstrations into sustained operation. His work focuses on building learning systems that allow robots to carry intent across time, space, and physical interaction, a capability that has historically separated research prototypes from truly usable machines.
“Autonomy fails when memory fails,” Canberk says. “If a system cannot carry its objective forward, capability does not matter.”
The task forces three problems to coexist in a single autonomous rollout: long-horizon planning, fine-grained dexterous manipulation, and room-scale navigation. None can be solved independently. Failure in any one collapses the entire chain. Treating this as a systems problem, rather than a demonstration, is what makes the work instructive for the broader field.
Long-Horizon Planning Without Resetting the World
Most robotic successes still operate within short temporal windows. Actions are executed, evaluated, and corrected within seconds. Household tasks do not work that way. They unfold over minutes, with compounding dependencies and no clean reset points.
“Real environments are adversarial to clean execution,” Canberk says. “The measure of autonomy is whether a system can maintain coherence as conditions drift.”
This is precisely where the table-to-dishwasher task constitutes a first-of-its-kind technical achievement. In a single autonomous rollout, the system sustains execution across 33 unique dexterous interactions, 68 total interaction events, and more than 130 feet of autonomous navigation, without resets, teleoperation, or task segmentation. Planning cannot be localized to a moment. Each decision commits the system to a future state it must continue to reason within.
Recent academic surveys underscore this gap. A 2025 research paper notes that long-horizon task execution remains one of the primary barriers preventing robots from operating autonomously in unstructured environments, despite advances in perception and control. The issue is not perception accuracy alone, but maintaining coherent intent over time.
By forcing the system to plan across dozens of interdependent actions: handling objects in a sensible order and navigating space with memory rather than reflex, the table-to-dishwasher task demonstrates an original contribution of major significance: it shows that long-horizon household autonomy can be achieved when planning is treated as a system-wide property rather than a sequence of local optimizations.
Dexterity as a First-Class Constraint
Manipulation has often been treated as a local problem. Grasp quality, force control, and finger placement are optimized in isolation. Household tasks collapse that abstraction. Dexterity becomes inseparable from planning.
“Treating manipulation as a bolt-on capability is a category error,” Canberk says. “In real environments, how an object is handled determines what the system can safely do next.”
In the table-to-dishwasher task, the robot must handle objects with drastically different physical properties: brittle glass, rigid ceramic, flexible packaging, and metallic utensils. Each interaction constrains the next. A poorly placed wine glass does not fail immediately; it fails later, when space runs out or force margins disappear.
This matters beyond a single task. According to the International Federation of Robotics’ 2025 service robotics outlook, failure modes in domestic robots are overwhelmingly tied to manipulation errors that compound over time rather than single-point mistakes. Reliability depends on how errors propagate, not whether they occur.
Framing dexterity this way shifts it from a motor-control problem to a systems-level design choice.
Navigation That Preserves Context
Navigation in robotics is often framed as a reactive control loop: perceive, move, correct. That framing works in constrained environments, but it breaks down in homes, where goals are distributed across rooms and frequently leave the robot’s field of view. In domestic settings, navigation is less about motion and more about maintaining intent while the environment changes.
In the table-to-dishwasher task, navigation cannot be isolated from the rest of the system. The robot must preserve spatial context while manipulating objects that alter future paths and constraints. Each movement between rooms depends on what is being carried, what has already been placed, and what remains unfinished. When spatial context is lost, recovery is not incremental; the task fails outright.
“Navigation only becomes meaningful when it is tied to purpose,” Canberk says. “A robot that can move efficiently but cannot remember why it is moving is not autonomous in any useful sense.”
This reframing exposes a broader limitation in many existing systems. Navigation stacks optimized for shortest paths or obstacle avoidance assume static goals and stable environments. Household tasks violate both assumptions. The robot’s own actions reshape the environment, and goals reappear only after long intervals, demanding continuity rather than reflex.
Why This Matters Beyond One Task
The table-to-dishwasher result does not claim that robots are ready for every home. It makes a narrower, more important claim: long-horizon autonomy is now a solvable engineering problem when treated as a unified system.
Industry momentum supports this framing. McKinsey’s 2025 outlook on AI-enabled robotics emphasizes that the next wave of value will come not from new skills, but from systems that can reliably chain existing skills under real-world constraints. Reliability, not novelty, is the bottleneck.
The implications extend beyond domestic robotics. Any environment that requires sustained autonomy—healthcare facilities, logistics hubs, or public infrastructure—faces the same structural challenges.
“What excites me is not one task,” Canberk concludes. “It is the idea that once continuity is solved, everything else compounds. Skills stop being demos and start becoming building blocks.”
The future of robotics will not be defined by isolated breakthroughs. It will be defined by whether autonomy can endure.
