RAMAMOORTHY, Subramanian, Collaborative Robotics, Edinburgh, June 10, 2016
<p class="MsoNormal" >also see private.</p>
<p>
RAMAMOORTHY, Subramanian, Collaborative Robotics, Edinburgh, June 10, 2016
(also see his YouTubes on collab robotics (soccer match) and using
Baxter for assistance)</p>
<p>
This is a noon seminar at Stanford. He is on sabbatical here for 6 months.</p>
<p>
robotics at Edinburgh...</p>
<p>
shows video of Freddy... a robot in Scotland Museum.</p>
<p>
from the 1960s.</p>
<p>
did pick and place task on objects on a moving table.</p>
<p>
and shows Shakey (SRI) famous robot from the 60s...</p>
<p>
what might be next...</p>
<p>
Collaborative Intelligent Robotics...</p>
<p>
coordinated interaction with others...</p>
<p>
shows collab robot from fraunhofer...</p>
<p>
intention aware collab with Baxter...</p>
<p>
an experiment from his lab...</p>
<p>
Baxter works an intel asst.</p>
<p>
baxter infers intention from workers gaze.</p>
<p>
"I think her needs the scissors"</p>
<p>
he's going to need the forceps….</p>
<p>
(I think of my appendectomy with the Filipina nurse)</p>
<p>
intention awareness...</p>
<p>
heeds the script but perhaps the surgeon changes his mind.</p>
<p>
from his lab in Edinburgh... Informatics Lab</p>
<p>
intention aware collab with the Baxter... get this video => AI/robo.</p>
<p>
driving forces and implications</p>
<p>
tasks can be ill-specified and environments are complex...</p>
<p>
interaction can help resolve incompleteness...</p>
<p>
and interaction variability and redundancy...</p>
<p>
hidden vars...</p>
<p>
</p>
<p>
Thinking fast and slow (like Daniel Kahnemann book)</p>
<p>
policies for dynamic and highly var. environs...</p>
<p>
1) can use prior experience and human demonstrations...</p>
<p>
and might there be a simpler template (like a script)</p>
<p>
so ... fast context-specific adaption of policies</p>
<p>
and slower... unsupervised learning of categories and concepts...</p>
<p>
and need to know properties of the system...</p>
<p>
eg are you working on a one gram part or a 150 kg engine.</p>
<p>
programming by demo (using Baxter)</p>
<p>
MIT... Rod Brooks company...</p>
<p>
easy to program by normal people.</p>
<p>
human just grabs the arm and moves it thru range of motion for the task (eg pick and place)</p>
<p>
Q... what is the task (inferring it)</p>
<p>
woman... wiping off dining room table.</p>
<p>
Manifold learning for generalizing motion...</p>
<p>
Havoutis IJRR 2013...</p>
<p>
robot carrying a tray like a waiter... despite being jostled.</p>
<p>
keeping tray stable and flat.</p>
<p>
representing motion: graphs probabilistic roadmap</p>
<p>
shows path planning... eg getting an arm into a tight engine space under the hood.</p>
<p>
not explicitly in the graph...</p>
<p>
topolog problem... do you go around object or take a direct path.</p>
<p>
he calls this homotopy...</p>
<p>
homology ... where are the holes in the 3d space...</p>
<p>
constructing filtration from samples</p>
<p>
where did the robot actually go...</p>
<p>
Cech complex...</p>
<p>
an abstract simplicial complex which is homotopy equivalent...</p>
<p>
and a Delaunay-Cech Complex...</p>
<p>
(obey properties like transitivity)</p>
<p>
can do trajectory classificaion in simplicial complex...</p>
<p>
is this path like this other one...</p>
<p>
shows a large filtration interval with only two cases (with a planar robot)</p>
<p>
either go to the R or the L around the solid object...</p>
<p>
gets samples in joint space in 4d.</p>
<p>
shows video of two different paths (that both go under an object...</p>
<p>
baxter expt in 3d with 97 trajectoires...</p>
<p>
can use hierarchical agglomerative clustering to combine classes.</p>
<p>
shows baxter grabbing a long cardboard cylinder...</p>
<p>
both arms go together to the L of the obstruction or L arm
goes to the L and R arm goes to the R of the obstruction...</p>
<p>
so... this is useful for planning...</p>
<p>
PR2 object handover...</p>
<p>
are objects on the red table or on the nearby blue table...</p>
<p>
want to make prediction...</p>
<p>
eg people moving toward the door...</p>
<p>
can coarsen the path by making increasingly large blobs around the path to the door.</p>
<p>
(he is on sabbatt. here for the year)</p>
<p>
Expt 1 qualitat behavior...</p>
<p>
eg indoor navigaiton...</p>
<p>
we building a particle filter (like a Kalman filter)</p>
<p>
coarse cloud and fine cloud (of trajecs)</p>
<p>
so can accept evidence from a fatter cloud of trajectories.</p>
<p>
suppose, someone says he's on the R side of the building...</p>
<p>
then that gives you evidence of a fat cloud of trajecs that go to the R.</p>
<p>
which bundle of paths is he in?</p>
<p>
and can run these at multiple scales of coarseness.</p>
<p>
Expt 2:</p>
<p>
understanding performance...</p>
<p>
MPF = particle filter...</p>
<p>
.. it helps...</p>
<p>
*</p>
<p>
decision making context...</p>
<p>
working together with unknown partners...</p>
<p>
with SV Albrect... see AIJ in 2016 eg.</p>
<p>
Ad Hoc coordination...</p>
<p>
robo interacts with person... but... is it gramma or a small child or roboticist.</p>
<p>
eg have a library of types (that could be FSM = fin state machines)</p>
<p>
stochastic bayesian game model...</p>
<p>
this is end player markov decn process.</p>
<p>
try to computer posterior beleif of the type</p>
<p>
(say have a library of types )</p>
<p>
but the true type space may be unknown... ie need to smapel fromt he true space)</p>
<p>
Bellman optical decn making...</p>
<p>
Harsanyi-Bellman Ad Hoc coord.</p>
<p>
eg consider a game with 2 or 3 players trying to get foods (with 5 vs 8 difft foods)</p>
<p>
did this at Royal Society summer science exhibition 201 2 in London...</p>
<p>
427 partic... from age 9 to age 72...</p>
<p>
human played against their agent...</p>
<p>
Prisoners dilemma</p>
<p>
vs rock paper scissors...</p>
<p>
we do much better in terms of total welfare
(ie combo of human and machine... get higher total score)</p>
<p>
but what if type space is wrong?</p>
<p>
summary </p>
<p>
policies for dynamic and high var enviro...</p>
<p>
and slow changing policy AND fast changing dynamics...</p>
<p>
so... difft time scales...</p>
<p>
and analysis of properties eg safety of a path.</p>
<p>
conclusion..</p>
<p>
collab robos bring up new opportunities.</p>
<p>
can also do programming by discussion... let use it to describe task</p>
<p>
and humans can help the robo get past barriers for learning...</p>
<p>
and abstract concepts that we wish to impart eg
like bubble wrap protects objects from breaking during a fall.</p>
<p>
*</p>
<p>
</p>
<p>
</p>