Low dimensional state representation learning with robotics priors in continuous action spaces

Nicolo Botteghi et al., University of Twente, IROS 2021

State representation learning

  • Three major categories
    1. Methods of encoding information to low-dimentional spaces by relying only on observation reconstruction. E.g. AE, VAE, denoising AE
      • Problem: ignoring small objects present in the observations, while these objects can be relevant for solving the task. Reconstruction are not useful, decoder may be not necessary.
      • Solution
    2. Model-based
      • Forward transition model, reward model, inverse model (?).
      • Problem: may collapse to trivial solutions, especially in case of sparse rewards
    3. All methods loosely constraining the state space using auxiliary loss functions injecting prior knowledge in the form of loss functions for training the encoder networks
      • Ref: Learning state representations with robotic priors, Autonomous Robots, 2015

Method

  • State and action
    • Agent’s state changes are directly related to the magnitude of the action taken.
    • Observation -> state prediction
    • State Use the magnitude of action connecting state prediction -> state distance
  • Simplicity prior 状态空间一定能缩小
  • Temporal coherence prior 时间上接近的state应该有较近的距离 应考虑到action带来的magnitude的影响
  • Proportionality prior 类似的action带来的state变化应类似
  • Repeatability prior 对于action,除了magnitude变化,还应考虑方向
  • Causality prior
  • 两级网络的结构
    • -> network structure REF: Low dimensional state representation learning with reward-shaped priors. the sensor modalities are independently processed by convolutional layers, flattened and merged through fully connected layers to create the final low-dimensional state prediction of dimension 5(?) dimension:
    • -> : mapping state to action

EXP

  • RGB and 2D LiDAR data merging.

Q: does this universal? to what range can be applied? Both architectures present three fully connected hidden layers of dimension 512?