CISC849:

Autonomous Robot Vision

​

http://goo.gl/0Im0yy

Prof. Christopher Rasmussen

​

February 11, 2016

Object detection/segmentation

Object recognition (from limited set)

Grasping: PR2 at TU Munich

(handle detection, affordance inference)

Hokuyo UTM-30LX ladar

• FOV: 270° H
• Range: 0.1-30 m
• Resolution: 0.25° / point (1080 points / scan)
• Update rate: 40 Hz, 43K points / s

​

Mechanism

Sample scan

MIT quadrotor doing obstacle avoidance

with onboard ladar sensing (start at 0:55)

3-D Point Clouds via Nodding/Spinning

• Tilt servo for household object manipulation (Jain & Kemp, ICRA 2010), uneven floor navigation for humanoid (Chestnutt et al., IROS 2009)

​

Chestnutt et al., IROS 2009

3-D Point Clouds via Nodding/Spinning

• Zebedee (Bosse et al., TRO 2012) Spring- mounted Hokuyo + IMU (e.g. Microstrain 3DM-GX3-25 OEM weighing 11.5 g)

CSIRO Zebedee demo (2:49)

3-D Point Clouds via Nodding/Spinning

• Off-axis spinning with slip ring (Singh et al., IROS 2011)

Velodyne HDL-64E ladar

• FOV: 360° H, 26.8° V
• Range: 0.9-120 m
• Resolution: 4167 x 64 (0.09° H, 0.33°-0.5° V) -> 267K points
• Update rate: 5-20 Hz (always 1.33 M points / s)
• Dimensions: 257 mm H x 224 mm W x 231 mm D
• Weight: 13.2 Kg
• Power: 12 V @ 4 A

​

Velodyne driving data (U. Koblenz-Landau)

Google Car (from IROS 2011 keynote)

(6:40-9:40) Note object clustering, multiple car tracking, stoplight detection, use of existing ladar map

DARPA Robotics Challenge (DRC)

http://theroboticschallenge.org (Trials Dec. 2013, Finals June 2015)

DRC: Why?

2013 DRC Trials: Scenarios

1. Drive a utility vehicle at the site
2. Travel dismounted across rubble
3. Remove debris blocking an entryway
4. Open a door and enter a building
5. Climb an industrial ladder and traverse an industrial walkway
6. Use a tool to break through a concrete panel
7. Locate and close a valve near a leaking pipe
8. Attach a hose

​

Robots at DRC

Schaft

BD

Atlas

CMU

Tartan Rescue

JPL

RoboSimian

DRC-Hubo

VT

Thor

NASA

Valkyrie

Schaft demonstrates events (1:54)

2015 DRC Finals

http://www.theroboticschallenge.org/overview, rules

Largely same tasks, but to be performed serially rather than individually

2015 DRC Finals: Top 5 Robots

1

2

3

4

5

Key Questions about Tasks

• How much is known about the environment, objects, lighting ahead of time?
• How much variability could there be?
• What robot skills will or could be engaged?
• Recognition (broad categories? specific objects?)
• Walking (flat ground? stairs? ramps? material?)
• Manipulation (precise grasping? pushing? hooking?)
• Bottom line: What do we need to understand about the scene (1) When robot first sees it, and (2) As robot performs the task?