Habitat 2.0 (Szot et al., 2021) ) is a simulation platform for training virtual robots in interactive 3D environments and complex physics-enabled scenarios. It includes a set of mobile manipulation tasks in which an agent controls a Fetch robot with a 7-DoF arm, mobile base and suction gripper to rearrange objects in apartment scenes.
We consider a challenging version of the Mobile-Pick task from Habitat 2.0, in which an agent must pick up a target object from a cluttered receptacle (e.g., a counter) while starting from a position in which the object is outside of the robot’s reach (thus, requiring navigation) without counting wih a dense goal specification sensor.
This benchmark requires a different environment because it uses a different version of habitat (0.2.3). To install the environment see the installation instructions.
Also make sure to specify your output folder for local runs and sweeps as well as your wandb entity to log results.
Run the following command to start training locally:
python run_rearrangement_vc.py model=vc1_vitb WANDB.run_name=<run_name> habitat_baselines.num_environments=6
Model can be vc1_vitb and vc1_vitl
You can use hydra to submit the job to a cluster by running:
python run_rearrangement_vc.py model=vc1_vitb WANDB.run_name=<run_name> habitat_baselines.num_environments=6 hydra/launcher=slurm_train_base -m
To evaluate a model run:
python run_rearrangement_vc.py model=vc1_vitb WANDB.run_name=<same_run_name_as_train> habitat_baselines.num_environments=10 hydra/launcher=slurm_eval_base RUN_TYPE=eval -m