A Contact Model based on Denoising Diffusion to Learn Variable Impedance Control for Contact-rich Manipulation

TL;DR

Stiffness tuning experiments conducted in simulated and real environments showed that the DCM achieved comparable performance to a conventional robot-based optimization method while reducing the number of robot trials.

Abstract

In this paper, a novel approach is proposed for learning robot control in contact-rich tasks such as wiping, by developing Diffusion Contact Model (DCM). Previous methods of learning such tasks relied on impedance control with time-varying stiffness tuning by performing Bayesian optimization by trial-and-error with robots. The proposed approach aims to reduce the cost of robot operation by predicting the robot contact trajectories from the variable stiffness inputs and using neural models. However, contact dynamics are inherently highly nonlinear, and their simulation requires iterative computations such as convex optimization. Moreover, approximating such computations by using finite-layer neural models is difficult. To overcome these limitations, the proposed DCM used the denoising diffusion models that could simulate the complex dynamics via iterative computations of multi-step denoising, thus improving the prediction accuracy. Stiffness tuning experiments conducted in simulated and real environments showed that the DCM achieved comparable performance to a conventional robot-based optimization method while reducing the number of robot trials.

Figures (16)

• Figure 1: The concept of learning of variable stiffness $\mathcal{K}$ formulated as a multi-objective optimization of the task objective$\mathcal{L}_{\mathrm{task}}$ and compliance objectives$\mathcal{L}_{\mathrm{comp}}$. (Top) Conventional framework (or robot-based optimization): the task objective is evaluated using a real robot johannsmeier2019frameworkwu2022prim8566177salehi2008impedancefateh2011adaptiveazimi2015stableokada2023learning. (Bottom) Proposed framework (or robot-free optimization): the task objective is evaluated using the proposed diffusion contact model (DCM) without robot trials.
• Figure 2: Illustration of Pareto solutions.
• Figure 3: Conceptual illustration of the diffusion contact model (DCM). (a) Block-diagram: DCM iteratively estimates the contact force trajectory $\mathcal{F}$ by $T$ steps of denoising, from the input containing the demonstrated trajectory $\tau^{\mathrm{demo}}$, reference trajectory $\tau^{\mathrm{attr}}$, and variable stiffness $\mathcal{K}$. (b/c) Examples of input/output, in which the orientation-related sequences are omitted for visibility. (d) Illustration of iterative denoising process.
• Figure 4: (a) Score function implementation by a state-of-the-art recurrent neural model, RetNet Sun2023-pp. Embed(i) denotes the embedding of the denoising step $i$ by the Gaussian Fourier projection song2020score. Tokenizer is realized by a fully-connected layer, which projects a concatenated input vector to the token $\mathbf{e}_{t}$. The number of dimensions of $\mathbf{e}_{t}$ was set as $128$. (b) Baseline model used in Sec. \ref{['sec:experiments']} which predicts forces in a single forward process (i.e., $T=1$). Bidirectional Gated Recurrent Units (Bi-GRU) were used previously gao2020learning; however RetNet was used here instread.
• Figure 5: Simulated wiping task environment and demonstrated trajectories. Demonstrations were performed by manually designed open-loop controllers. (a) Trajectories for training: we demonstrated 80 skills of spiral motion with different bowl sizes and directions, and then performed Bayesian optimization with $N=20$ trials to reproduce each demonstration, which yielded 1,600 trajectories. (b) Trajectories for test: we demonstrated a single skill of wavelike motion, and performed Bayesian optimization with $N=30$ trials.