Team: Bruce Kim, Indraneel Patil Bruce and I moved in together and were discussing buying a robot vacuum, but then decided to build one instead! Some design principles: Buy as many off the shelf parts as possible Keep the budget less than 500$ It should have a minimum battery life to be able to navigate around the house and vacuum the floor, so that we only have to charge it once a week Hardware Software Since we dont have any compute on the robot to do inference we decided to stream image frames from the robot to a laptop and then run inference on the robot and return result back to the robot for navigation. We teleoperated the robot around the house to collect tuples of image frame and a discrete navigation action. Our discrete navigation action set was: COMMANDS = { keyboard.Key.up: "FORWARD\n", keyboard.Key.down : "REVERSE\n", keyboard.Key.left: "TURN_CCW\n", keyboard.Key.right: "TURN_CW\n", keyboard.Key.space: "STOP\n" } Then we trained a simple CNN with behavior cloning to learn from these expert trajectories. Observations: Lots of false positive reversing, like for these there was no need to reverse so maybe it doesnt have a good perception of depth? -> It doesnt need a good perception of depth it only needs to worry about the ground plane in front of it In the kitchen seems to just keep oscillating between forward and reverse -> This could be lack of good data Should have created a validation / training split It kept getting STOP for this, where clearly there is space ahead for it to move into Kept getting FORWARD for this, when clearly an obstacle can be seen Dont think it is trained super well There is some motion blur but most images are really nice STOP action can get the robot stuck in front of an obstacle, is that really needed? Can we remove STOP? This paper only use the STOP action when robot has reached the goal I feel like what its learnt is quite random but it has learnt to back up when very close to an obstacle or its about to hit something which...
First seen: 2026-04-08 09:19
Last seen: 2026-04-08 10:20