BC-RNN & BC-Transformer

Behavior Cloning with Recurrent Neural Networks (BC-RNN) is a behavior cloning model that uses a recurrent neural network (RNN) to encode temporal dependencies in demonstration data. The model learns to map observations (e.g., images, sensor data) to actions by imitating expert demonstrations. Behavioral Cloning with an Transformer network (BC-Transformer) is a behavioral cloning model share the similar architecture with BC-RNN but replace the RNN backbone with a Transformer backbone.

Model Architecture:

• The BC-RNN/BC-Transformer model consists of:

  • Observation Encoder:

    • Processes high-dimensional observations (e.g., images) into a compact representation.

  • Policy Network:

    • A recurrent (e.g., LSTM or GRU) or transformer neural network that models temporal dependencies in the demonstration data.

    • Predicts actions based on the encoded observations and task embeddings.

  • Task Embedding Module:

    • Encodes task descriptions (e.g., natural language instructions) into a latent space.

    • Conditions the RNN policy on the task context.

More Information:

• Full paper: https://arxiv.org/abs/2108.03298

• Homepage: https://robomimic.github.io/

• Github link: ARISE-Initiative/robomimic

Model Conversion