Interview with Jean Pierre Sleiman, creator of the paper “Versatile multicontact planning and management for legged loco-manipulation”

Image from paper “Versatile multicontact planning and management for legged loco-manipulation“. © American Affiliation for the Development of Science

We had the possibility to interview Jean Pierre Sleiman, creator of the paper “Versatile multicontact planning and management for legged loco-manipulation”, lately revealed in Science Robotics.

What’s the subject of the analysis in your paper?
The analysis subject focuses on growing a model-based planning and management structure that permits legged cell manipulators to deal with various loco-manipulation issues (i.e., manipulation issues inherently involving a locomotion aspect). Our examine particularly focused duties that may require a number of contact interactions to be solved, moderately than pick-and-place functions. To make sure our method shouldn’t be restricted to simulation environments, we utilized it to unravel real-world duties with a legged system consisting of the quadrupedal platform ANYmal outfitted with DynaArm, a custom-built 6-DoF robotic arm.

Might you inform us concerning the implications of your analysis and why it’s an fascinating space for examine?
The analysis was pushed by the need to make such robots, particularly legged cell manipulators, able to fixing quite a lot of real-world duties, reminiscent of traversing doorways, opening/closing dishwashers, manipulating valves in an industrial setting, and so forth. A typical method would have been to deal with every process individually and independently by dedicating a considerable quantity of engineering effort to handcraft the specified behaviors:

That is sometimes achieved by way of the usage of hard-coded state-machines by which the designer specifies a sequence of sub-goals (e.g., grasp the door deal with, open the door to a desired angle, maintain the door with one of many ft, transfer the arm to the opposite aspect of the door, cross by way of the door whereas closing it, and so forth.). Alternatively, a human knowledgeable might show the right way to remedy the duty by teleoperating the robotic, recording its movement, and having the robotic be taught to imitate the recorded conduct.

Nevertheless, this course of may be very sluggish, tedious, and vulnerable to engineering design errors. To keep away from this burden for each new process, the analysis opted for a extra structured method within the type of a single planner that may mechanically uncover the mandatory behaviors for a variety of loco-manipulation duties, with out requiring any detailed steerage for any of them.

Might you clarify your methodology?
The important thing perception underlying our methodology was that all the loco-manipulation duties that we aimed to unravel may be modeled as Process and Movement Planning (TAMP) issues. TAMP is a well-established framework that has been primarily used to unravel sequential manipulation issues the place the robotic already possesses a set of primitive expertise (e.g., choose object, place object, transfer to object, throw object, and so forth.), however nonetheless has to correctly combine them to unravel extra complicated long-horizon duties.

This attitude enabled us to plan a single bi-level optimization formulation that may embody all our duties, and exploit domain-specific information, moderately than task-specific information. By combining this with the well-established strengths of various planning strategies (trajectory optimization, knowledgeable graph search, and sampling-based planning), we had been in a position to obtain an efficient search technique that solves the optimization drawback.

The primary technical novelty in our work lies within the Offline Multi-Contact Planning Module, depicted in Module B of Determine 1 within the paper. Its general setup may be summarized as follows: Ranging from a user-defined set of robotic end-effectors (e.g., entrance left foot, entrance proper foot, gripper, and so forth.) and object affordances (these describe the place the robotic can work together with the thing), a discrete state that captures the mix of all contact pairings is launched. Given a begin and purpose state (e.g., the robotic ought to find yourself behind the door), the multi-contact planner then solves a single-query drawback by incrementally rising a tree by way of a bi-level search over possible contact modes collectively with steady robot-object trajectories. The ensuing plan is enhanced with a single long-horizon trajectory optimization over the found contact sequence.

What had been your primary findings?
We discovered that our planning framework was in a position to quickly uncover complicated multi- contact plans for various loco-manipulation duties, regardless of having supplied it with minimal steerage. For instance, for the door-traversal state of affairs, we specify the door affordances (i.e., the deal with, again floor, and entrance floor), and solely present a sparse goal by merely asking the robotic to finish up behind the door. Moreover, we discovered that the generated behaviors are bodily constant and may be reliably executed with an actual legged cell manipulator.

What additional work are you planning on this space?
We see the offered framework as a stepping stone towards growing a completely autonomous loco-manipulation pipeline. Nevertheless, we see some limitations that we purpose to handle in future work. These limitations are primarily related to the task-execution part, the place monitoring behaviors generated on the idea of pre-modeled environments is barely viable underneath the idea of a fairly correct description, which isn’t all the time easy to outline.

Robustness to modeling mismatches may be tremendously improved by complementing our planner with data-driven strategies, reminiscent of deep reinforcement studying (DRL). So one fascinating course for future work could be to information the coaching of a strong DRL coverage utilizing dependable knowledgeable demonstrations that may be quickly generated by our loco-manipulation planner to unravel a set of difficult duties with minimal reward-engineering.

In regards to the creator

Jean-Pierre Sleiman acquired the B.E. diploma in mechanical engineering from the American College of Beirut (AUB), Lebanon, in 2016, and the M.S. diploma in automation and management from Politecnico Di Milano, Italy, in 2018. He’s at present a Ph.D. candidate on the Robotic Programs Lab (RSL), ETH Zurich, Switzerland. His present analysis pursuits embody optimization-based planning and management for legged cell manipulation.