Revision history [back]

We don't see that many robot RL papers that use ABB YuMi and that makes us wonder whether there are limitations/challenges/etc. when it comes to using YuMi for research.

There is/was no support for EGM on YuMi officially. While it is possible reach somewhat decent on-line performance using just RAPID (with fi something like kth-ros-pkg/yumi), EGM is really the future and in my opinion the only control interface that would make something like RL on actual hw interesting.

It's most likely not the only reason you don't see many papers, but for researchers interested in running experiments on actual hw (which is a requirement for quite some venues now) and/or looking to run RL training phases not in simulation but on real hw, not having a decent, real-time external control interface will make them look at other robots. And there are quite a few (just not humanoid style upper torsos with two arms).

ABB has released two packages that help with integrating EGM (into any application, not just ROS): ros-industrial/abb_libegm and its companion ros-industrial/abb_librws. These are officially supported by ABB and under active development.

Sawyer has a very nice ROS interface, which YuMi does not (yet) have. That is certainly a negative on the side of YuMi: it will take you somewhat longer to get started with YuMi. However, as @stevemacenski already wrote: Rethink is no more, and it's somewhat unclear (to me at least) what the future of the research side of Baxter and Sawyer is.

Community-wise I would expect there to be a bit more people using Sawyer with ROS than YuMi (Sawyer was released not that long before Rethink stopped operations), but EGM is used with ABB's other robots, so there is a decent number of people using that.

And just to avoid a false dilemma: there are quite a few other robot brands that have decent, real-time capable external motion control interfaces. You may have already evaluated those, and crossed them off your list, but just in case you may want to look at other options as well.

We don't see that many robot RL papers that use ABB YuMi and that makes us wonder whether there are limitations/challenges/etc. when it comes to using YuMi for research.

There is/was no support for EGM on YuMi officially. While it is possible reach somewhat decent on-line performance using just RAPID (with fi something like kth-ros-pkg/yumi), EGM is really the future and in my opinion the only control interface that would make something like RL on actual hw interesting.interesting (in the case of ABB robots).

It's most likely not the only reason you don't see many papers, but for researchers interested in running experiments on actual hw (which is a requirement for quite some venues now) and/or looking to run RL training phases not in simulation but on real hw, not having a decent, real-time external control interface will make them look at other robots. And there are quite a few (just not humanoid style upper torsos with two arms).

ABB has released two packages that help with integrating EGM (into any application, not just ROS): ros-industrial/abb_libegm and its companion ros-industrial/abb_librws. These are officially supported by ABB and under active development.

Sawyer has a very nice ROS interface, which YuMi does not (yet) have. That is certainly a negative on the side of YuMi: it will take you somewhat longer to get started with YuMi. However, as @stevemacenski already wrote: Rethink is no more, and it's somewhat unclear (to me at least) what the future of the research side of Baxter and Sawyer is.

Community-wise I would expect there to be a bit more people using Sawyer with ROS than YuMi (Sawyer was released not that long before Rethink stopped operations), but EGM is used with ABB's other robots, so there is a decent number of people using that.

And just to avoid a false dilemma: there are quite a few other robot brands that have decent, real-time capable external motion control interfaces. You may have already evaluated those, and crossed them off your list, but just in case you may want to look at other options as well.

We don't see that many robot RL papers that use ABB YuMi and that makes us wonder whether there are limitations/challenges/etc. when it comes to using YuMi for research.

There is/was was no support for EGM on YuMi officially. That has changed recently, with the introduction of RobotWare 6.09. While it is possible reach somewhat decent on-line performance using just RAPID (with fi something like kth-ros-pkg/yumi), EGM is really the future and in my opinion the only control interface that would make something like RL on actual hw interesting (in the case of ABB robots).

It's most likely not the only reason you don't see many papers, but for researchers interested in running experiments on actual hw (which is a requirement for quite some venues now) and/or looking to run RL training phases not in simulation but on real hw, not having a decent, real-time external control interface will make them look at other robots. And there are quite a few (just not humanoid style upper torsos with two arms).

ABB has released two packages that help with integrating EGM (into any application, not just ROS): ros-industrial/abb_libegm and its companion ros-industrial/abb_librws. These are officially supported by ABB and under active development.

Sawyer has a very nice ROS interface, which YuMi does not (yet) have. That is certainly a negative on the side of YuMi: it will take you somewhat longer to get started with YuMi. However, as @stevemacenski already wrote: Rethink is no more, and it's somewhat unclear (to me at least) what the future of the research side of Baxter and Sawyer is.

Community-wise I would expect there to be a bit more people using Sawyer with ROS than YuMi (Sawyer was released not that long before Rethink stopped operations), but EGM is used with ABB's other robots, so there is a decent number of people using that.

And just to avoid a false dilemma: there are quite a few other robot brands that have decent, real-time capable external motion control interfaces. You may have already evaluated those, and crossed them off your list, but just in case you may want to look at other options as well.

We don't see that many robot RL papers that use ABB YuMi and that makes us wonder whether there are limitations/challenges/etc. when it comes to using YuMi for research.

There was no support for EGM on YuMi officially. That has changed recently, with the introduction of RobotWare 6.09. While it is possible reach to achieve somewhat decent on-line performance using just RAPID (with fi something like kth-ros-pkg/yumi), EGM is really the future and in my opinion the only control interface that would make something like RL on actual hw interesting (in the case of ABB robots).

It's most likely not the only reason you don't see many papers, but for researchers interested in running experiments on actual hw (which is a requirement for quite some venues now) and/or looking to run RL training phases not in simulation but on real hw, not having a decent, real-time external control interface will make them look at other robots. And there are quite a few (just not humanoid style upper torsos with two arms).

ABB has released two packages that help with integrating EGM (into any application, not just ROS): ros-industrial/abb_libegm and its companion ros-industrial/abb_librws. These are officially supported by ABB and under active development.

Sawyer has a very nice ROS interface, which YuMi does not (yet) have. That is certainly a negative on the side of YuMi: it will take you somewhat longer to get started with YuMi. However, as @stevemacenski already wrote: Rethink is no more, and it's somewhat unclear (to me at least) what the future of the research side of Baxter and Sawyer is.

Community-wise I would expect there to be a bit more people using Sawyer with ROS than YuMi (Sawyer was released not that long before Rethink stopped operations), but EGM is used with ABB's other robots, so there is a decent number of people using that.

And just to avoid a false dilemma: there are quite a few other robot brands that have decent, real-time capable external motion control interfaces. You may have already evaluated those, and crossed them off your list, but just in case you may want to look at other options as well.