See, feel, act: Hierarchical learning for complex manipulation skills with multisensory fusion

Research Topic

Inspired by:

• Humans are able to seamlessly integrate tactile and visual stimuli with their intuitions to explore and execute complex manipulation skills.

While learning contact-rich manipulation skills, we face two important challenges: active perception and hybrid behavior.

怎样理解active preception 和 hybrid behavior是本文关键。

Methods

Evaluation metric

In this study, they evaluated the robot’s ability to play the game by counting the number of successful consecutive block extractions in randomly generated towers.

需要注意的一点是：
This paper emphasizes physics modeling and does not explicitly evaluate the adversarial nature of the game.

Task specifications

• sensing
  The robot have access to its own pose, the pose of the blocks, and the forces applied to it at every time step.
• action primitives
  The robot uses two ‘primitive’ actions, push and extract/place.
• base exploration policy
  The robot has access to a base exploration policy for data collection.
• termination criteria
  1. all blocks have been explored
  2. a block is dropped outside the tower
  3. the tower has toppled
• tower and robot specifications

Simulation

使用MuJoCo作为仿真环境。在仿真环境中比较以下几种approach(前三种是model，后一种是policy)的performance:

1. HMAs
   这篇文章提出的hierarchical model abstractions
2. NN
   a feed-forward neural network as a representative nonhierarchical model-based approach
3. MOR
   a mixture of regressions model as a generic hierarchical model-based approach
4. PPO
   implementation of RL as a model-free approach

All methods have access to the same set of states, actions, and MPC.