Embedded Files
Gregory Kahn, Pieter Abbeel, Sergey Levine
Berkeley AI Research (BAIR), University of California, Berkeley
paper | github
Motivation
Motivation
Recent technological advancements have spurred the development of autonomous mobile robots, from sidewalk delivery robots and agricultural inspection quadcopters to autonomous vehicles.
One of the primary metrics for measuring progress has been how far can the robot travel before the robot fails and a human must intervene?
Below, we see a sidewalk delivery robot being tested. A person is monitoring the robot, and disengages and repositions the robot each time it fails.
Typically, these failures are then used to help the autonomy developers debug the software. However, this debugging process is highly nontrivial, system dependent, and laborious, especially for learning-based components.
Idea
Idea
We investigate how to use these disengagements as a direct learning signal for navigation.
LaND: Learning to Navigate from Disengagements
LaND: Learning to Navigate from Disengagements
LaND learns to navigate by:
- Leveraging existing testing datasets
consisting of only the robot’s onboard sensors---such as camera images, the robot’s commanded actions---such as the steering angle, and whether the robot autonomy mode was engaged or disengaged.
2. Training a disengagement prediction model
2. Training a disengagement prediction model
that takes as input the current sensor observations and a sequence of future planned actions, and predicts whether the robot will be engaged or disengaged in the future.
3. Using this disengagement prediction model for planning and control
3. Using this disengagement prediction model for planning and control
so that the robot can plan and execute actions that avoid disengagements.
LaND is then able to
LaND is then able to
navigate near a parked bicycle, a sharp turn, near dense foliage, and in sun glare
is better able to navigate on 2.3 km of never-before-seen sidewalks compared to prior imitation learning and reinforcement learning methods
and improve as it more data is gathered
Page updated
Google Sites
Report abuse