Res-HIL starts with a policy trained from demonstrations and keeps it frozen.
Reinforcement learning then trains a residual policy to correct its actions.
During training, a human operator can take control when the robot is likely to fail.
Those interventions provide examples of the corrections to make and help identify
autonomous actions that led to trouble.
-
Residual actions. The base policy produces an action from the robot's observation.
The residual policy uses the observation and base action to produce a correction.
The robot executes the sum of the two actions. We initialize the residual policy
to output zero, so training starts with the base policy's behavior.
-
Human interventions. The operator can override the robot's action and guide it
back to a recoverable state. We use the human action to teach the residual policy
what correction to make. We also penalize autonomous actions immediately before
an intervention.
-
Two replay buffers. The online buffer stores both autonomous and human-controlled
transitions. A second buffer stores the initial demonstrations and intervention
transitions. Each training batch draws equally from the two buffers.
-
Continuous learning. While the robot interacts with the task, a separate
learning loop samples transitions, updates the critics and residual policy,
and sends the updated weights back to the robot.
-
Learning the correction. We train the residual policy with TD3 using the
robot's task reward and the intervention-shaped penalties. A behavior cloning
loss teaches it to match corrections from demonstrations and interventions,
while a regularization term discourages unnecessarily large corrections.