[Paper Review] Time Series Motion Generation Considering Long Short-Term Motion

Kazuki Fujimoto, Sho Sakaino+1 more|arXiv (Cornell University)|Sep 9, 2019

Robot Manipulation and Learning28 references4 citations

TL;DR

This paper proposes a novel recurrent neural network (RNN) model, the PLT model, that separately handles long-term (low-frequency) and short-term (high-frequency) motion by adjusting input sampling periods for position and force data. By delaying position input and using multi-decimation-inspired frequency separation, the model enables robust letter writing on unlearned paper heights, outperforming standard RNNs in generalization and stability during long-horizon tasks.

ABSTRACT

Various adaptive abilities are required for robots interacting with humans in daily life. It is difficult to design adaptive algorithms manually; however, by using end-to-end machine learning, labor can be saved during the design process. In our previous research, a task requiring force adjustment was achieved through imitation learning that considered position and force information using a four-channel bilateral control. Unfortunately, tasks that include long-term (slow) motion are still challenging. Furthermore, during system identification, there is a method known as the multi-decimation (MD) identification method. It separates lower and higher frequencies, and then identifies the parameters characterized at each frequency. Therefore, we proposed utilizing machine learning to take advantage of the MD method to infer short-term and long-term (high and low frequency, respectively) motion. In this paper, long-term motion tasks such as writing a letter using a pen fixed on a robot are discussed. We found differences in suitable sampling periods between position and force information. The validity of the proposed method was then experimentally verified, showing the importance of long-term inference with adequate sampling periods.

Motivation & Objective

• To address the challenge of generating long-term (slow) motion in robot manipulation tasks, especially those requiring force and position coordination.
• To improve generalization in imitation learning by incorporating long-term motion dynamics, which standard RNNs fail to capture.
• To overcome limitations in existing models that cannot extrapolate to unlearned environmental conditions like varying paper heights.
• To explore the use of system identification principles—specifically the multi-decimation (MD) method—in machine learning for motion generation.
• To validate that appropriate sampling periods for position and force data are critical for stable and accurate long-term motion inference.

Proposed method

• The PLT model uses two separate RNN branches: one for long-term position inference with delayed input sampling, and one for short-term force control with high-frequency sampling.
• The model is inspired by the multi-decimation (MD) system identification method, which separates low- and high-frequency dynamics for parameter identification.
• Training data is collected via bilateral control, where an operator writes on paper at three heights (10 mm, 40 mm, 70 mm), and the robot learns the master-slave response mapping.
• The model learns to infer the master robot’s motion from the slave robot’s response, with position and force signals processed at different effective sampling rates.
• The PLT model explicitly delays position input to allow for long-term trajectory planning, reducing prediction-induced vibrations.
• A comparison is made with a standard CONV. RNN model that processes data at a fixed 20 ms interval, highlighting the advantage of adaptive sampling.

Experimental results

Research questions

• RQ1 Can a deep learning model effectively learn and generalize long-term motion trajectories in robot manipulation tasks, such as writing on paper at varying heights?
• RQ2 How does the choice of sampling period for position and force inputs affect the stability and accuracy of motion generation in imitation learning?
• RQ3 Can the multi-decimation (MD) system identification method be effectively adapted to improve long-term motion inference in neural networks?
• RQ4 Does delaying position input improve the robustness and generalization of RNN-based motion generation models in unlearned environments?
• RQ5 Why do standard RNNs fail to generalize to unlearned paper heights, while the PLT model succeeds?

Key findings

• The PLT model successfully wrote the letter 'A' on paper at unlearned heights (25 mm, 55 mm, and 85 mm), demonstrating strong generalization beyond training data.
• The standard CONV. RNN model failed entirely to write the letter at any of the unlearned heights, highlighting its inability to handle long-term motion dynamics.
• The PLT model maintained consistent writing pressure across different paper heights, as shown by stable torque responses in Fig. 11, indicating effective force regulation.
• At 100 mm height, the PLT model could no longer write clearly, indicating a limit to its extrapolation capability despite strong generalization at lower heights.
• The model reduced vibration in sequential position predictions by delaying position input, improving motion smoothness and stability.
• The MD-inspired model structure enabled effective separation of long-term (trajectory) and short-term (force) dynamics, proving critical for complex motion tasks.