Subgoal Diffuser: Coarse-to-fine Subgoal Generation to Guide Model Predictive Control for Robot Manipulation

TL;DR

A diffusion-based method is presented that guides an MPC method to accomplish long-horizon manipulation tasks by dynamically specifying sequences of subgoals for the MPC to follow, and it improves the planning performance of an MPC method, and also outperforms prior diffusion-based methods.

Abstract

Manipulation of articulated and deformable objects can be difficult due to their compliant and under-actuated nature. Unexpected disturbances can cause the object to deviate from a predicted state, making it necessary to use Model-Predictive Control (MPC) methods to plan motion. However, these methods need a short planning horizon to be practical. Thus, MPC is ill-suited for long-horizon manipulation tasks due to local minima. In this paper, we present a diffusion-based method that guides an MPC method to accomplish long-horizon manipulation tasks by dynamically specifying sequences of subgoals for the MPC to follow. Our method, called Subgoal Diffuser, generates subgoals in a coarse-to-fine manner, producing sparse subgoals when the task is easily accomplished by MPC and more dense subgoals when the MPC method needs more guidance. The density of subgoals is determined dynamically based on a learned estimate of reachability, and subgoals are distributed to focus on challenging parts of the task. We evaluate our method on two robot manipulation tasks and find it improves the planning performance of an MPC method, and also outperforms prior diffusion-based methods.

Figures (4)

• Figure 1: Middle: Our system consists of a diffusion model that generates subgoals in a coarse-to-fine manner, and a low-level MPC controller that tracks the subgoals. The diffusion model generates subgoals recursively until all subgoals are reachable from their predecessors. Left: To estimate the reachability, we learn a function that estimates the number of steps required to move between the two subgoals. If the prediction is smaller than the horizon of the MPC, we assume it is reachable. Right: Our Subgoal Diffuser is conditioned on current state, goal state, subgoals from the previous level, and (optionally) an SDF of the environment. The subgoals from the previous level will be redistributed so that they are equally spaced in terms of execution steps.
• Figure 2: Prediction of finer subgoals $\boldsymbol{\tau}_{\mathcal{G}}^{l}$ are generated contitioned on the coarse subgoals $\boldsymbol{\tau}_{\mathcal{G}}^{l-1}$. To compute the conditioning $\hat{\boldsymbol{\tau}}_{\mathcal{G}}^{l}$, we encode $\boldsymbol{\tau}_{\mathcal{G}}^{l-1}$ into latent space and upsample it using linear interpolation. Top: Without redistribution, the new subgoals are evenly distributed, ignoring the relative distance between consecutive subgoals. Thus, the subgoals that are far apart remain unreachable. Bottom: $\boldsymbol{\tau}_{\mathcal{G}}^{l-1}$ are redistributed in latent space using the estimated pairwise distance. By doing so, more subgoals will be filled in to the unreachable segment.
• Figure 3: Number of subgoals as the robot progresses.
• Figure 4: Snapshots of rope and notebook manipulation tasks in simulation (top two rows) and the real world (bottom two rows). The subgoals are visualized in yellow for simulation experiments. More visualizations can be found on our https://sites.google.com/view/subgoal-diffuser-mpc/home