Home

Habitat challenge (paper)

Current plan: Recreate the winner solution from the last year Train and deploy in the environment Deploy in the real world Optimize performance

Last year’s winner for object-nav (what we are implementing):

Arnold (active neural slam; this guy transferred his model to the real world) Paper Active Neural Slam (github)

3 major parts of the project:

• Reinforcement Learning
• Computer Vision
• Robotics

Kamal’s Notes here

Steps:

• Gather 5-10 papers Reinforcement Learning SpinningUp, Sutton&Barto RL Book Hierarchical learning link1 Proximal Policy Optimisation Learning Exploration Policies for Navigation, here slam and exploration Partially Observable Markov Decision Process (POMDP) link1
• Read those papers
• go through the code ANS
• run inference with their pre-trained model

RL https://arxiv.org/abs/1708.05866 https://arxiv.org/abs/1811.12560

Is an option of using a Deterministic local policy actually using heuristic local policy? They mention using ULP instead of training

Goal(s)?

Reimplement the paper Code from scratch Will need to train the model in the cloud Research how much training capability would be needed. Use their hyperparameters In tensorflow (?) Build an affordable robot and deploy the model there Find a cheap locobot alternative (under $300?) Find parts (already have jetson nano) Figure out if we can run this on jetson nano