Whole-body manipulation is a strength of humans but a weakness of robots. The robot interprets each possible contact point between the box and the carrier’s fingers, arms, or torso as a separate contact event. This task becomes difficult to prepare for as soon as one considers the billions of possible contact events. Now, MIT researchers can streamline this technique, called contact-rich manipulation planning. An artificial intelligence approach called smoothing is used to reduce the number of judgments needed to find a good manipulation plan for the robot from the vast number of contact occurrences.
New developments in RL have demonstrated amazing results in manipulating through contact-rich dynamics, something that was previously challenging to achieve using model-based techniques. While these techniques were effective, it has yet to be known why they succeeded while model-based approaches failed. The overarching objective is to grasp and make sense of these factors from a model-based vantage point. Based on these understandings, scientists work to merge RL’s empirical success with the models’ generalizability and efficacy.
The hybrid nature of contact dynamics presents the greatest challenge to planning through touch from a model-based perspective. Since the ensuing dynamics are non-smooth, the Taylor approximation is no longer valid locally, and the linear model built using the gradient quickly breaks down. Since both iterative gradient-based optimization and sampling-based planning use local distance metrics, the local model’s invalidity poses serious difficulties for both. In response to these problems, numerous publications have attempted to take contact modes into account by either listing them or providing examples of them. These planners, who have a model-based understanding of the dynamic modes, often switch between continuous-state planning in the current contact mode and a discrete search for the next mode, leading to trajectories with a few-mode shifts here and there.
The first thing researchers have added is proof that the two smoothing strategies are theoretically equivalent for basic systems under the framework. In addition, using this framework, the authors demonstrate how to efficiently compute the locally linear models (i.e., gradients) of the smoothed dynamics in real-time, and they demonstrate that the qualitative characteristics and empirical performance of the two smoothing schemes are comparable across various complex examples.
What is the best model for gradient-based contact-rich manipulation planning? It is at the center of the model-based planning approach taken by researchers. To efficiently compute local linearizations for planning, they believe this model must be (i) numerically robust and (ii) differentiable. For the planner to see far into the future with minimal effort, the model must be able to (iii) forecast long-term behavior. Finally, the model needs to (iv) be smooth to provide more information gradients across contact modes.
The second improvement is a complete model of contact dynamics. In particular, they suggest an implicit time-stepping contact model that is convex. Anitescu’s relaxation of frictional contact limits leads to convexity. However, it does introduce some mildly non-physical behavior in reality. Compared to the standard Linear Complementarity Problem (LCP) formulation, convexity offers significant numerical benefits.
The quasi-dynamic assumption is commonly employed in robotic manipulation because it allows long-term predictability. There is no need for variables representing velocity or damping in quasi-dynamic models because kinetic energy is lost at each time step. They verify and test the quasi-dynamic contact model by simulating and executing the same input paths in Drake, a high-fidelity second-order simulator on hardware. If the system under consideration is highly damped and dominated by frictional forces, the results suggest that the model can better approximate the second-order dynamics.
In addition, a log-barrier relaxation can be used to soften the contact model analytically. As is typical in the interior point method for convex systems, a log-barrier function is utilized to enforce the hard contact restrictions in this relaxation strategy flexibly. Further work demonstrates that the implicit function theorem provides a straightforward method for calculating the gradients of the smoothed contact model. Finally, experts believe RL’s aim of performing global optimization with stochasticity is another major element behind its empirical success. Nonlinear dynamical planning using deterministic models typically yields non-convex optimization problems, where the quality of many local minima might be decisive.
The last contribution addresses this shortcoming by integrating RRT’s global search capabilities with those of smoothing-based contact mode abstraction. Using a novel distance measure derived from the local smoothed models, researchers have made it possible for RRT to search through the limits imposed by contact dynamics.
Overall Contributions
Scientists determine the qualitative and empirical equivalence of randomized and analytic smoothing techniques on straightforward systems.
They show contact-rich manipulation planning can benefit greatly from a convex, differentiable formulation of quasi-dynamic contact dynamics and associated analytic smoothing.
Researchers integrate contact mode smoothing with sampling-based motion planning to achieve effective global planning via highly rich contact dynamics, filling a gap in the spectrum of existing approaches.
Researchers clarify the mathematical meaning of smoothing a function and several strategies for computing its local approximations before discussing contact in complicated systems. Their goal is to present a unified picture of smoothing techniques and the relationships between them.
The researchers were inspired to do this by the striking difference between the success of RL in empirical situations with lots of human contact and the failure of model-based approaches. They have shown that traditional model-based approaches can effectively tackle planning for contact-rich manipulation by identifying the pitfalls in the existing model-based methods for planning, understanding how RL was able to alleviate such traps, and resolving them with model-based techniques. By enabling efficient online planning in the order of a minute and being generalizable with respect to environments and tasks, the contribution offers a powerful alternative to existing tools in RL that rely on heavy offline computation on the order of hours or days. They review some of the factors that made this possible.
In a nutshell, they were inspired to do this study after realizing the dramatic gap between the success of RL in empirical contexts and the struggle of model-based approaches to this problem. They have shown that traditional model-based approaches can effectively tackle planning for contact-rich manipulation by identifying the pitfalls in the existing model-based methods for planning, understanding how RL was able to alleviate such pitfalls, and resolving them with model-based techniques. By enabling efficient online planning in the order of a minute and being generalizable with respect to environments and tasks, the contribution offers a powerful alternative to existing tools in RL that rely on heavy offline computation on the order of hours or days. They review some of the factors that made this possible.
Initially identified as a flaw in model-based approaches, the need to explicitly enumerate and assess modes has been mitigated by RL’s stochastic smoothing. Next, they’ve brought up another flaw in model-based techniques: second-order transients might cause short-sighted linearizations that don’t help with long-term strategy. They have proposed the Convex Differentiable Quasi-Dynamic Contact (CQDC) model to address this shortcoming. They have demonstrated the usefulness of the touch model through numerous theoretical arguments and experiments. They also demonstrated that the contact dynamics can be relaxed analytically with a log barrier by first evaluating the model’s structure. They conducted studies demonstrating the computing advantages of analytic smoothing over randomized smoothing.
In conclusion, they found that smoothing-based model-based strategies have been linked to local trajectory optimization. Compared to RL-based techniques that try to do global search, they have proven less successful in challenging issues due to their susceptibility to local minima. However, SBMP techniques for contact-rich systems have avoided the trap of mode enumeration by explicitly taking into account contact modes. The work contributes by closing a gap in pre-existing approaches by fusing mode smoothing with RRT, wherein the exploration phase of RRT was guided by a local approximation to the smooth surrogate based on the local Mahalanobis metric. By combining these three advancements they have made it possible for model-based and RL-based approaches to achieve efficient global motion planning for very contact-rich and high-dimensional systems. In the future, they will employ a highly streamlined planner version to drive policy searches or perform real-time motion planning. They anticipate that this enhancement will allow robots to locate contact-rich designs online in previously unexplored areas within seconds of planning time.
n