RMG: Real-Time Expressive Motion Generation with Self-collision Avoidance for 6-DOF Companion Robotic Arms

TL;DR

This work presents RMG, a real-time expressive motion generation framework for 6-DOF robotic arms that leverages a dance-derived expressive motion dataset, diffusion-based trajectory generation in both joint and Cartesian spaces, and a PSO-based collision-avoidance optimization. By transforming human dance data into robot motions, training joint-space and Cartesian-space diffusion models, and post-processing with a collision-free optimizer, the approach achieves expressive, smooth trajectories under real-time constraints (under roughly 0.5 seconds). The joint-space model outperforms the Cartesian-space model in planning accuracy, expressivity, and robustness, and the system is demonstrated on a low-cost Mycobot 280 with a gesture-based interaction scenario. Limitations include overshoot and speed-control challenges, with future work aiming to endow the robot with autonomous emotional reasoning to select endpoints and durations contextually.

Abstract

The six-degree-of-freedom (6-DOF) robotic arm has gained widespread application in human-coexisting environments. While previous research has predominantly focused on functional motion generation, the critical aspect of expressive motion in human-robot interaction remains largely unexplored. This paper presents a novel real-time motion generation planner that enhances interactivity by creating expressive robotic motions between arbitrary start and end states within predefined time constraints. Our approach involves three key contributions: first, we develop a mapping algorithm to construct an expressive motion dataset derived from human dance movements; second, we train motion generation models in both Cartesian and joint spaces using this dataset; third, we introduce an optimization algorithm that guarantees smooth, collision-free motion while maintaining the intended expressive style. Experimental results demonstrate the effectiveness of our method, which can generate expressive and generalized motions in under 0.5 seconds while satisfying all specified constraints.

Figures (7)

• Figure 1: Overview of RMG. Blue arrows indicate the robotic dataset construction pipeline, while red arrows represent the robotic motion generation process.
• Figure 2: Motion transforming Process. After filtering and coordinate aligning the human motion trajectory, the total velocity at Points 20 and 21 decide the left or right trajectory. The chosen trajectory is then scaled and mapped into the robot arm's joint space. Finally, motion optimization generates a smooth, collision-free, and feasible trajectory.
• Figure 3: Motion Diffusion Generation Overview. The J Model and C Model share the same model structure. This system learns to denoise motion sequences from time $t = T$ to $t = 0$, where $t$ represents the diffusion steps. $l_{valid}$ is the valid sequence length to determine the motion time. In the C Model, start state $x_{\text{start}}$ and end state $x_{\text{end}}$ denote coordinate values, while in the J Model, they correspond to $\theta_1$, $\theta_2$, $\theta_3$. All input conditions are converted into vector feature $x \in \mathbb{R}^{b \times h}$, and the noise sequence is converted to $x_{\text{motion}} \in \mathbb{R}^{b \times l \times h}$, where $L$ is the whole sequence length and $h$ is set to 256 as the hidden layer dimension. $N$ refers to the stack number. The postprocessing of the denoised sequence can be found in Sec. \ref{['sec:processing']}.
• Figure 4: Feature Fusion Module. In the $w$ sequence, the gray areas indicate invalid parts, and the color intensity in the valid sections reflects the weight. The sum of the two sequences gives the valid length weight.
• Figure 5: Point C trjectory $t_C$ in the J Model