Whole-body manipulation is a strength of humans but a weakness of robots. MIT researchers have developed a technique called contact-rich manipulation planning to address this weakness. This technique uses an artificial intelligence approach called smoothing to reduce the number of judgments needed for the robot to find a good manipulation plan. It has been challenging to achieve manipulation through contact-rich dynamics using previous model-based techniques, but new developments in reinforcement learning (RL) have shown promising results.
The hybrid nature of contact dynamics presents a challenge for planning through touch from a model-based perspective. Local models built using the gradient quickly break down because the ensuing dynamics are non-smooth. To overcome these difficulties, researchers have attempted to incorporate contact modes into the planning process. However, these planners often switch between continuous-state and discrete search modes, resulting in trajectories with only a few mode shifts.
To address these challenges, researchers have made two important improvements. First, they have shown that two smoothing strategies are equivalent for basic systems. They have also developed a complete model of contact dynamics using an implicit time-stepping contact model that is convex. This convex model offers significant numerical benefits compared to traditional linear complementarity problem formulations. The quasi-dynamic assumption allows for long-term predictability in robotic manipulation.
Researchers have also integrated contact mode smoothing with sampling-based motion planning to achieve effective global planning in contact-rich environments. By combining these techniques, they have filled a gap in existing approaches and provided a powerful alternative to RL tools that require heavy offline computation.
The researchers have demonstrated that traditional model-based approaches can effectively tackle planning for contact-rich manipulation by understanding the limitations of existing methods and leveraging the strengths of RL. Their contributions have been supported by theoretical arguments and experiments, showing the usefulness and advantages of their approach.
In conclusion, the researchers have addressed the weaknesses of robots in whole-body manipulation by developing a contact-rich manipulation planning technique. They have integrated RL and model-based techniques to achieve effective global planning in contact-rich environments. Their work has demonstrated the potential of smoothing-based strategies and the benefits of a convex, differentiable formulation of contact dynamics.