Compare robotic foundation models.

This is the same reference used in Technical mode. Expand it whenever you need the exact numbers, assumptions or further explanation.

Start with the question

A policy follows a new instruction on its training robot. Does that establish transfer to a different arm?

Mechanism and worked comparison

Robot evidence depends on the body as well as the model. Sensors, joint limits, control frequency, contact and demonstration collection affect what an action means. ACT studies imitation with action chunks on a bimanual setup. Diffusion Policy represents conditional action sequences through denoising. RT-2 investigates transfer from vision-language learning into token-based robot action prediction. The cards link their 2023 primary publications or project pages.

A common comparison asks where demonstrations came from, which observations enter the policy, how actions are represented, how often feedback changes a plan, and which complete tasks were measured. A chunk may smooth behavior while delaying reaction. A multimodal generator may represent alternative actions without deciding which is safe. Language-related transfer may help a semantic choice while leaving grounding, permission and physical feasibility unresolved.

Do not rank these papers from isolated success percentages on different tasks or embodiments. Published failure cases and omitted conditions matter alongside successes. Write a constrained claim such as “this comparison supports the reported task behavior under the studied setup,” then propose a matched test on a new object or sensor condition. The browser cards support research literacy; they do not run the paper models, authenticate a reproduction or qualify a learner to deploy a robot.

Try and explain

Compare ACT, Diffusion Policy and RT-2 with the common checklist. Identify a possible failure that the model representation itself cannot prevent.

Open the reproducible study project or use the implementation studio.