search

askforalfred / alfred

ALFRED - A Benchmark for Interpreting Grounded Instructions for Everyday Tasks
MIT License
352 stars 77 forks source link

planner pickup action #128

Closed RishiHazra closed 1 year ago

RishiHazra commented 1 year ago

Hi,

We are working on this dataset to learn operators/actions from it. It seems like some of the conditions asserted by the ground truth PDDL operators used for generating the dataset are not being strictly obeyed. As an example, even though the ground truth pickup operators have these preconditions: (atLocation ?a ?l) (objectAtLocation ?o ?l) the object is only actually at the same location as the agent in 6% of the pickup transitions.

For instance: here the agent and the knife are not in the same location

(atLocation agent1 loc_bar_5_bar_8_bar_0_bar_45)
(objectAtLocation Knife_bar__plus_01_dot_17_bar__plus_00_dot_79_bar__plus_02_dot_71 loc_bar_4_bar_7_bar_0_bar_45)
(= (distance loc_bar_4_bar_7_bar_0_bar_45 loc_bar_5_bar_8_bar_0_bar_45) 7)

However, the agent is still able to pick up the knife as a part of the traj['plan']['high_pddl']. Just wondering if there's a minumum distance threshold that the ff-planner considers? Any help is appreciated, thanks !

RishiHazra commented 1 year ago

It was problem with AI2-THOR internal state tracking and not a planner issue. The plan gets executed correctly, however the state metadata returned was noisy.