Generalization
New Colors
New Colors
Agent trained on manipulating 3 cubes of different colors (red, green and blue).
During inference, the agent generalizes to a larger number of cubes in colors not seen in RL training (yellow and purple).
Goal Image - Training
Agent Rollout - Training
Agent Rollout - Inference
Goal Image - Inference
State-Goal Mismatch
State-Goal Mismatch
Agent trained on manipulating 3 cubes of different colors.
During inference, the environment configuration is different than the goal  (blue cube is not present in the environment).
The agent is not "distracted" by the irrelevant goal object.
Goal Image
Agent Rollout