Efficient Robotic Object Search via HIEM: Hierarchical Policy Learning with Intrinsic-Extrinsic Modeling

This is the source code for our HIEM framework and the baseline methods we mentioned in the paper.

paper | video

Requirements

Our code is developed and tested under the following dependencies:

• python==2.7.15
• scipy==1.2.0
• numpy==1.15.4
• tensorflow==1.6.0
• tf Slim
• opencv==3.2.0-dev

Before running the code, please specify the path to the code directory in the config.json.

Download our pre-processed data sourced from House3D and extract here for our robotic object search task.

Download our pre-trained models and put them in the corresponding code directories for training and/or evaluating our method.

Training

To train our model HIEM in the paper, run this command:

# Specify the parameters in HIEM/train.sh, 
# and from HIEM/
./train.sh

or

# From HIEM/
python train.py \
    --load_model=True \
    --default_scenes=<enviroments_to_train> \
    --default_targets=<target_objects_to_train> \
    --pretrained_model_path=${PATH_TO_PRETRAINED_MODEL} \
    --model_path=${PATH_TO_MODEL} 

where the pretrained_model_path is ../h-DQN/result*_mt_for_pretrain/model and the environments and their target objects for training can be found in readme.

To train other baseline methods mentioned in the paper, run the same command from the corresponding directories.

Evaluation and Results

To evaluate our method HIEM for the robotic object search task on House3D,

• run the command,

  # From HIEM/
  CUDA_VISIBLE_DEVICES=-1 python evaluate.py \
  --max_episodes=1 \
  --load_model=True \
  --model_path="result1_mt_pretrain/model" \
  --evaluate_file='../random_method/1s6t_1.txt' \
  --default_scenes='5cf0e1e9493994e483e985c436b9d3bc' \
  --default_targets='music' \
  --default_targets='television' \
  --default_targets='table' \
  --default_targets='stand' \
  --default_targets='dressing_table' \
  --default_targets='heater' \ 

  to reproduce the results of our method on the environment 1 as follows,

To evaluate other environments and target objects, change model_path, evaluate_file, default_scenes and default_targets accordingly.

To evaluate other baseline methods, run the same command from the corresponding directories.