SAKI turns separate human demonstrations into longer mobile robot tasks. It is still a lab result.
A University of Sydney team reports a method for reusing human-video interactions in new layouts and combining them into tasks such as tidying and wiping, without showing a home product.
Real-robot experiments demonstrate skill reuse across layouts and the composition of independently demonstrated interactions into continuous mobile tasks, including tidying and wiping.Lu, Zhao and Zhi, SAKI paper
The result
A team at the University of Sydney has posted SAKI, short for Skill Assembly and Kinematic Imitation. The research asks a useful household-robot question: can a robot take separate demonstrations of ordinary interactions and join them into a longer job when the room is not arranged exactly like the original video?
The authors report real-robot experiments that reuse skills across layouts and combine independently demonstrated interactions into continuous mobile tasks, including tidying and wiping. Their project page shows recorded task sequences involving putting a can in a bin, pouring from a paper cup, wiping a whiteboard, placing objects and opening a box.
That is a research claim, not a product announcement. The paper was submitted to arXiv on September 28, 2026, and does not announce a consumer robot, price, release date or home deployment.
How it works
SAKI starts from a human video, but it does not try to copy every movement. The paper says it extracts object-centric skill assets: the parts that matter to the interaction, such as contact, clearance, orientation and the intended final relationship between an object and its destination. The robot then generates its own path for the room in front of it.
For a simple example, a demonstration of placing an object needs the object to arrive in the right place and orientation. It does not require a robot to take the same detour a person took around a table. SAKI separates those task requirements from motion that can change with the robot's starting position and the current layout.
The system then selects, orders and binds the skills to objects in the present scene. It carries scene estimates and the robot's ending configuration from one skill to the next. During execution, visual feedback can revise the remaining plan when the estimated position of an object changes.
The project page describes recorded trials in changed layouts and longer sequences built from independently demonstrated interactions. It also lists examples where the robot maintains contact through a task, including wiping, opening a door and holding a lid while placing an object.
What it means for a home robot
The practical idea is sound. A useful home robot cannot be trained only for one kitchen arranged one way. It needs to preserve the point of a task while changing its route around a different chair, countertop or starting position. It also needs to keep track of what happened earlier in a multi-step job.
SAKI addresses those two pieces: reuse across layouts and continuity between task segments. The reported demonstrations are closer to a tidy-up routine than a full household chore. Still, they are more relevant to home help than a benchmark that moves a single object on a fixed tabletop.
The work also highlights why a short robot video can be misleading. Completing a wipe, a placement or a pour separately is different from completing several actions in a room where objects and viewpoints change. The paper's contribution is the machinery for connecting those actions, not a claim that the robot has mastered general domestic work.
Caveats
The evidence is limited to a research system and recorded real-robot trials. The paper says its experiments include tidying and wiping, but it does not establish all-day reliability, safe operation around people, recovery from failures or performance in occupied homes. The project page presents nine recorded trials for reuse across scenes and several recorded sequences for longer tasks; those are demonstrations, not evidence of broad deployment.
SAKI also depends on supplied goals and task dependencies. That matters. Choosing what should be cleaned, finding every needed object and deciding when a task is complete remain separate problems from composing motion once the job is defined.
For now, SAKI is a useful research result about a narrow but stubborn problem: making demonstrations reusable when the world changes. It is not evidence that a consumer humanoid is ready to take over the house.
We labeled this story Research — A lab result, not a product. How we label claims