Tsinghua and AGIBOT: choosing the data that will teach a task to the robot
Chinese programs / Industry and mobility / English edition
What you will learn to do
Define a distinct set of learning situations and a set of tests for a robotic task.
Original VOCE course. Chinese benchmark: data collection via teleoperation in the Tsinghua / AGIBOT curriculum. The case presented is fictional.
The course
In the Tsinghua / AGIBOT curriculum, students collect teleoperation data before training and deployment. This step places task experience at the center of system development.
A fictional workshop films one hundred grasping gestures performed on the same part, always placed in the same orientation. The team has a volume of data. They do not yet know if this volume represents the task they wish to entrust to the robot.
The manager must describe the expected variations. Parts may change position. Presentation conditions may differ between two teams. Certain situations must be rejected rather than processed.
Collection must be organized around this scope. Repeating an easy case one hundred times is no substitute for observing different conditions. It is also necessary to keep track of the situations collected and the interventions required.
The test set serves to examine the system's behavior once developed. If the same situations were used to build and tune the system, the result may be too favorable toward what has already been encountered. The company must plan for a separation appropriate to its task and its data.
Finally, the decision concerns the uncovered conditions. The manager must not conclude that the robot knows how to handle a category of objects simply because it resembles that of the tests. They can limit the initial scope and plan for verifications before each extension.
Your decision
The team has collected 80 easy grasps and 20 grasps requiring human intervention. They want to delete these 20 sequences and reserve 10 easy sequences for testing. What information do you request before accepting this plan?
Formulate an answer before opening the answer key.
Open the reasoned answer key
Ask why the interventions were necessary and whether the corresponding situations will be part of the actual work. Deleting them without examination may mask a central difficulty. Some sequences may be unusable for training while remaining relevant for designing the tests.
The test set must represent the conditions of the announced scope, not just the easy cases. Document the data separation and the absent categories. The answer key does not prescribe a universal training method.
Production to keep
A collection plan with covered variations, excluded data, and situations reserved for testing.
Evaluation criterion
The participant examines the coverage of situations and the separation of tests, beyond the number of sequences.
Sources and references
- Tsinghua, laboratory and AGIBOT course, September 14, 2026
- Tsinghua iCenter, course on embodied intelligence scenarios, June 3, 2026
- NIST, AI RMF 1.0, 2023
Original VOCE creation, published exclusively on Academy. Chinese programs are documentary references; no university partnership is claimed.