# US 2026/0091493 A1 · Machines learning assembly tasks using pre-trained skill libraries

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2026/0091493 A1
- Document ID: US20260091493A1
- Status: Published application
- Kind: A1
- Application number: 19/271,620
- Filed: 2025-07-16
- Published: 2026-04-02
- Earliest claimed: 2024-09-30
- Assignees: NVIDIA Corporation
- Applicants: NVIDIA Corporation
- Inventors: Yijie Guo; Bingjie Tang; Iretiayo Akinola; Dieter Fox; Abhishek Gupta; Yashraj Narang
- Web page: https://patentgrep.com/patent/US20260091493A1

## CPC

- B25J 9/1661: {characterised by task planning, object-oriented languages}
- B25J 9/163: {learning, adaptive, model based, rule based expert control}
- B25J 9/1664: {characterised by motion, path, trajectory planning}
- B25J: MANIPULATORS; CHAMBERS PROVIDED WITH MANIPULATION DEVICES ({manipulators specially adapted for use in surgery A61B34/70; manipulators used in cleaning hollow articles B08B9/04}; manipulators associated with rolling mills B21B39/20; manipulators associated with forging machines B21J13/10; {manipulators associated with picking-up and placing mechanisms B23P19/007}; means for holding wheels or parts thereof B60B30/00; {vehicles with ground-engaging propulsion means, e.g. walking members B62D57/02, B62D57/032; devices for picking-up and depositing articles or materials between conveyors B65G47/90, B65G47/91; manipulators with gripping or holding means for transferring packages B65H67/065}; cranes B66C; {manipulators used in the protection or supervision of pipe-line installations F17D5/00; walking equipment adapted for nuclear steam-generators F22B37/006}; manipulators specially adapted for, or associated with, nuclear reactors G21C; {apparatus used for handling wafers during manufacture or treatment of semiconductor H10P72/50})

## Abstract

To determine the actions of a machine assembly system (e.g., robotic assembly system), a relevant skill policy can be selected from a library of skill policies. The target task can be used to determine target task information, specifically, the geometry of one or more objects (e.g., parts) of the assembly task, one or more dynamic parameters of the objects, and expert actions that are specified. These parameters can be used as input into the skill library to predict the most relevant skill policy. The target task and the source task (e.g., a potential skill policy) can be compared on the metrics of geometry, dynamics, and expert actions. The best match skill policy can be returned as the relevant skill policy. This policy can then be modified, such as fine-tuning, to improve the real-world assembly outcomes.

## Claims

### Claim 1

A tool for skill retrieval and skill adaptation (SRSA), comprising:
an interface configured to receive data, wherein the data includes target task information associated with a target task; and
one or more processors to perform one or more operations that include:
retrieving, using the target task information, a relevant skill policy from a skill library to perform the target task, wherein the retrieving is based on a predicted success of the relevant skill policy and the relevant skill policy includes a source geometry of at least one source object and at least one task trajectory of the at least one source object, where the at least one source object is matched to a target object of the target task, and
modifying the relevant skill policy for the target task using the target task information.

### Claim 2 (depends on 1)

The tool as recited in claim 1, wherein the target task is an assembly task.

### Claim 3 (depends on 1)

The tool as recited in claim 1, wherein the retrieving utilizes a task feature learning and a transfer success prediction, wherein the task feature learning comprises:
extracting target geometry features from point cloud reconstruction of the target object involved in the target task;
extracting dynamics features by predicting a next state of the target object given a current state and an action of the target object; and
extracting expert action features by predicting the action from an observed state transition using inverse dynamics prediction.

### Claim 4 (depends on 3)

The tool as recited in claim 3, where the transfer success prediction is performed using a neural network trained to predict a transferability between a source task of the source object and the target task, wherein the neural network receives as input a combination of source task features and target task features, and outputs a predicted transfer success score representing how well the relevant skill policy applies to the target task, and where the source task features include of the at least one source object encoding of the source geometry, encoding of task dynamics, and encoding of expert actions, and the target task features include of the target object encoding of the target geometry, encoding of task dynamics, and encoding of expert actions.

### Claim 5 (depends on 3)

The tool as recited in claim 3, wherein retrieving the relevant skill policy includes computing the transfer success prediction between each source policy in the skill library and the target task, and selecting the source policy with a highest predicted transfer success score as the relevant skill policy for the target task.

### Claim 6 (depends on 5)

The tool as recited in claim 5, wherein the retrieved relevant skill policy further utilizes fine-tuning for the target task.

### Claim 7 (depends on 6)

The tool as recited in claim 6, wherein the fine-tuning utilizes a proximal policy optimization.

### Claim 8 (depends on 1)

The tool as recited in claim 1, wherein the predicted success is determined using a transfer success predictor trained with labels obtained from applying source policies to the target task, and wherein the transfer success predictor outputs a scalar score representing how well a source policy applies to the target task.

### Claim 9 (depends on 1)

The tool as recited in claim 1, wherein the skill library is generated using previous tasks at a previous time interval.

### Claim 10 (depends on 1)

The tool as recited in claim 1, wherein the relevant skill policy is retrieved using a combination of object geometries of the at least one source object, object dynamics of the at least one source object, and expert actions on the at least one source object to represent a previous task in the skill library, and the combination is compared to the target task information.

### Claim 11 (depends on 1)

The tool as recited in claim 1, wherein the modifying the relevant skill policy uses a self-imitation learning process.

### Claim 12 (depends on 1)

The tool as recited in claim 1, wherein the one or more operations are performed within a simulated environment and the modified relevant skill policy is stored in the skill library as a previous task.

### Claim 13 (depends on 1)

The tool as recited in claim 1, wherein the one or more operations further include directing one or more robots using the modified relevant skill policy to perform the target task.

### Claim 14 (depends on 1)

The tool as recited in claim 1, wherein at least one skill policy in the skill library includes disassembly trajectory information enabling representation learning of dynamics and expert actions.

### Claim 15 (depends on 14)

The tool as recited in claim 14, wherein the disassembly trajectory information is stored with disassembly actions, representing expert actions for task feature learning.

### Claim 16

A robotic assembly system, comprising:
a skill library configured to store skill policies, wherein each of the skill policies include source task information that includes a source task, a source geometry of at least one source object of the source task, at least one dynamic of the at least one source object, and at least one expert action of the at least one source object; and
one or more processors configured to receive a target task and target task information of the target task, retrieve a relevant skill policy from the skill policies to perform the target task utilizing a predicted success of the relevant skill policy to perform the target task with the target task information, and modify the relevant skill policy for the target task using the target task information, wherein the target task information includes a target geometry of at least one target object of the target task and at least one dynamic of the at least one target object.

### Claim 17 (depends on 16)

The robotic assembly system as recited in claim 16, further comprising:
a robot configured to perform the target task using at least the target object and the modified relevant skill policy.

### Claim 18 (depends on 16)

The robotic assembly system as recited in claim 16, wherein the target task is an assembly task of two objects, one of which is the target object.

### Claim 19 (depends on 16)

The robotic assembly system as recited in claim 16, further comprising:
a machine learning system configured to work with the one or more processors to analyze the target task information and to select the relevant skill policy using a machine learning process of the machine learning system.

### Claim 20 (depends on 16)

The robotic assembly system as recited in claim 16, wherein the one or more processors are part of or executing on a central processor unit (CPU) or a graphics processor unit (GPU).

### Claim 21 (depends on 16)

The robotic assembly system as recited in claim 16, wherein the modified relevant skill policy is used by a robotic planner or a robotic controller system to direct the movement of a robotic assembler.

### Claim 22 (depends on 16)

The robotic assembly system as recited in claim 16, wherein the relevant skill policy includes the source geometry of at least one source object.

### Claim 23 (depends on 16)

The robotic assembly system as recited in claim 16, wherein the relevant skill policy includes at least one task trajectory of the at least one source object.

### Claim 24

A method of operating a robot for a new assembly task, comprising:
receiving a target task, wherein the target task is an assembly task of at least two objects;
determining target task information associated with the target task;
retrieving, using the target task information, a relevant skill policy from a skill library to perform the target task, wherein the retrieving is based on a predicted success of the relevant skill policy; and
modifying the relevant skill policy for the target task.

### Claim 25 (depends on 24)

The method as recited in claim 24, further comprising:
directing operations of a robotic assembly system using the modified relevant skill policy.

### Claim 26 (depends on 24)

The method as recited in claim 24, further comprising:
applying the modified relevant skill policy to a set of previously unseen target tasks with unknown object geometries, wherein the modified relevant skill policy enables zero-shot assembly in real-world environments.

### Claim 27 (depends on 24)

The method as recited in claim 24, further comprising:
using the modified relevant skill policy by a robotic planner or a robotic controller system.

### Claim 28 (depends on 24)

The method as recited in claim 24, further comprising:
using the modified relevant skill policy as an input for a machine learning process or a robotic learning process.

### Claim 29 (depends on 24)

The method as recited in claim 24, further comprising:
simulating a robotic assembly system, the target task, and the skill library within a machine learning system.

### Claim 30

A non-transitory computer program product having a series of operating instructions stored on a non-transitory computer-readable medium that directs a robotic assembly system when executed thereby to perform operations, the operations comprising:
receiving a target task, wherein the target task is an assembly task of at least two objects;
determining target task information associated with the target task;
retrieving, using the target task information, a relevant skill policy from a skill library to perform the target task, wherein the retrieving is based on a predicted success of the relevant skill policy; and
modifying the relevant skill policy for the target task.

### Claim 31 (depends on 30)

The non-transitory computer program product recited in claim 30, wherein the retrieving further comprises:
utilizing task feature learning and transfer success prediction to select the target task.

### Claim 32 (depends on 31)

The non-transitory computer program product recited in claim 31, wherein the task feature learning includes capturing geometry features from point cloud reconstruction, capturing dynamics features from next state prediction, and capturing expert action features from inverse dynamics prediction.

### Claim 33 (depends on 31)

The non-transitory computer program product recited in claim 31, wherein the transfer success prediction is performed using a neural network trained to predict a success of applying a potential skill policy to the target task using input features from a source task and the target task, including at least one of encoded geometry, dynamics, or expert actions, and the potential skill policy with a highest predicted transfer success is selected as the relevant skill policy.
