Single-view robot pose and joint angle estimation via render & compare

RoboPose. (a) Given a single RGB image of a known articulated robot in an unknown configuration (left), RoboPose estimates the joint angles and the 6D camera-to-robot pose (rigid translation and rotation) providing the complete state of the robot within the 3D scene, here illustrated by overlaying the articulated CAD model of the robot over the input image (right). (b) When the joint angles are known at test-time (e.g. from internal measurements of the robot), RoboPose can use them as an additional input to estimate the 6D camera-to-robot pose to enable, for example, visually guided manipulation without fiducial markers.

Citation

If you use this code in your research, please cite the paper:

@inproceedings{labbe2021robopose,
title= {Single-view robot pose and joint angle estimation via render & compare}
author={Y. {Labb\'e} and J. {Carpentier} and M. {Aubry} and J. {Sivic}},
booktitle={Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR)},
year={2021}}

Table of content

• Overview
• Installation
• Downloading and preparing data
• Note on GPU parallelization
• Reproducing results
  • Quantitative and qualitative evaluation
  • Running inference
  • Training models

Overview

This repository contains the code for the full RoboPose approach and for reproducing all the results from the paper (training, inference and evaluation).

Installation

git clone --recurse-submodules https://github.com/ylabbe/robopose.git
cd robopose
conda env create -n robopose --file environment.yaml
conda activate robopose
python setup.py install
mkdir local_data

The installation may take some time as several packages must be downloaded and installed/compiled. If you plan to change the code, run python setup.py develop.

Downloading and preparing data

All data used (datasets, models, results, ...) are stored in a directory local_data at the root of the repository. Create it with mkdir local_data or use a symlink if you want the data to be stored at a different place. We provide the utility robopose/scripts/download.py for downloading required data and models. All of the files can also be downloaded manually.

Robot URDF & CAD models

python -m robopose.scripts.download --robot=owi
python -m robopose.scripts.download --robot=kuka
python -m robopose.scripts.download --robot=panda
python -m robopose.scripts.download --robot=baxter

DREAM & CRAVES Datasets

python -m robopose.scripts.download --datasets=craves.test
python -m robopose.scripts.download --datasets=dream.test

# Only for re-training the models
python -m robopose.scripts.download --datasets=craves.train
python -m robopose.scripts.download --datasets=dream.train

Pre-trained models

python -m robopose.scripts.download --model=panda-known_angles
python -m robopose.scripts.download --model=panda-predict_angles
python -m robopose.scripts.download --model=kuka-known_angles
python -m robopose.scripts.download --model=kuka-predict_angles
python -m robopose.scripts.download --model=baxter-known_angles
python -m robopose.scripts.download --model=baxter-predict_angles
python -m robopose.scripts.download --model=owi-predict_angles

DREAM & CRAVES original results

python -m robopose.scripts.download --dream_paper_results
python -m robopose.scripts.download --craves_paper_results

Notes:

• Dream results were extracted using the official code from https://github.com/NVlabs/DREAM.
• CRAVES results were extracted using the code provided with the paper. We slightly modified this code to compute the errors on the whole LAB dataset, the code used can be found on our fork.

Note on GPU parallelization

Training and evaluation code can be parallelized across multiple gpus and multiple machines using vanilla torch.distributed. This is done by simply starting multiple processes with the same arguments and assigning each process to a specific GPU via CUDA_VISIBLE_DEVICES. To run the processes on a local machine or on a SLUMR cluster, we use our own utility job-runner but other similar tools such as dask-jobqueue or submitit could be used. We provide instructions for single-node multi-gpu training, and for multi-gpu multi-node training on a SLURM cluster.

Single gpu on a single node

# CUDA ID of GPU you want to use
export CUDA_VISIBLE_DEVICES=0
python -m robopose.scripts.example_multigpu

where scripts.example_multigpu can be replaced by scripts.run_pose_training or scripts.run_robopose_eval (see below for usage of training/evaluation scripts).

Configuration of job-runner for multi-gpu usage

Change the path to the code directory, anaconda location and specify a temporary directory for storing job logs by modifying `job-runner-config.yaml'. If you have access to a SLURM cluster, specify the name of the queue, it's specifications (number of GPUs/CPUs per node) and the flags you typically use in a slurm script. Once you are done, run:

runjob-config job-runner-config.yaml

Multi-gpu on a single node

# CUDA IDS of GPUs you want to use
export CUDA_VISIBLE_DEVICES=0,1
runjob --ngpus=2 --queue=local python -m robopose.scripts.example_multigpu

The logs of the first process will be printed. You can check the logs of the other processes in the job directory.

On a SLURM cluster

runjob --ngpus=8 --queue=gpu_p1  python -m robopose.scripts.example_multigpu

Reproducing results using pre-trained models

We provide the inference results on all datasets to reproduce the results from the paper. You can download these results, generate the tables and qualitative visualization of our predictions on the test datasets. The results will be downloaded to local_data/results.

Downloading inference results

# Table 1, DREAM paper results (converted from the original format)
python -m robopose.scripts.download --results=dream-paper-all-models

# Table 1, DREAM Known joint angles
python -m robopose.scripts.download --results=dream-known-angles

# Table 1, DREAM Unknown joint angles
python -m robopose.scripts.download --results=dream-unknown-angles

# Table 2, Iterative results
python -m robopose.scripts.download --results=panda-orb-known-angles-iterative

# Table 3, Craves-Lab
python -m robopose.scripts.download --results=craves-lab

# Table 4, Craves Youtube
python -m robopose.scripts.download --results=craves-youtube

# Table 5, Analysis of the choice of reference point
python -m robopose.scripts.download --results=panda-reference-point-ablation

# Table 6, Analysis of the choice of the anchor part
python -m robopose.scripts.download --results=panda-anchor-ablation

# Sup. Mat analysis of the number of iterations
python -m robopose.scripts.download --results=panda-train_iterations-ablation

You can generate the numbers from the tables from these inference/evaluation results using the notebook notebooks/generate_results.ipynb.

You can generate visualization of the results using the notebook notebooks/visualize_predictions.ipynb.

Running inference

We provide the code for running inference and re-generate all results. This is done using the run_robot_eval script. The results were obtained using the following commands:

## Main results and comparisons
# DREAM datasets,  DREAM models
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=dream-panda  --model=dream-all-models --id 1804
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=dream-baxter --model=dream-all-models --id 1804
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=dream-kuka  --model=dream-all-models --id 1804

# DREAM datasets, ours (known joints)
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=dream-panda  --model=knownq --id 1804
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=dream-baxter --model=knownq --id 1804
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=dream-kuka   --model=knownq --id 1804

# DREAM datasets, ours (unknown joints)
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=dream-panda  --model=unknownq --id 1804
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=dream-baxter --model=unknownq --id 1804
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=dream-kuka   --model=unknownq --id 1804

# CRAVES LAB dataset
runjob --ngpus=8 python scripts/run_robot_eval.py --datasets=craves-lab --model=unknownq --id 1804

# CRAVES Youtube dataset
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=craves-youtube --model=unknownq-focal=500 --id 1804
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=craves-youtube --model=unknownq-focal=750 --id 1804
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=craves-youtube --model=unknownq-focal=1000 --id 1804
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=craves-youtube --model=unknownq-focal=1250 --id 1804
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=craves-youtube --model=unknownq-focal=1500 --id 1804
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=craves-youtube --model=unknownq-focal=1750 --id 1804
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=craves-youtube --model=unknownq-focal=2000 --id 1804
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=craves-youtube --model=unknownq-focal=5000 --id 1804


## Ablations
# Online evaluation, Table 2
runjob --ngpus=8 python scripts/run_robot_eval.py --datasets=dream-panda-orb --model=knownq --id 1804 --eval_all_iter
runjob --ngpus=1 python scripts/run_robot_eval.py --datasets=dream-panda-orb --model=knownq-online --id 1804

# Analysis of reference point, Table 5
python -m robopose.scripts.download --models=ablation_reference_point
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=dream-panda-orb  --model=knownq-link0 --id 1804
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=dream-panda-orb  --model=knownq-link1 --id 1804
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=dream-panda-orb  --model=knownq-link5 --id 1804
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=dream-panda-orb  --model=knownq-link2 --id 1804
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=dream-panda-orb  --model=knownq-link4 --id 1804
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=dream-panda-orb  --model=knownq-link9 --id 1804

# Analysis of anchor part, Table 6
python -m robopose.scripts.download --models=ablation_anchor
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=dream-panda-orb  --model=unknownq-link1 --id 1804
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=dream-panda-orb  --model=unknownq-link2 --id 1804
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=dream-panda-orb  --model=unknownq-link5 --id 1804
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=dream-panda-orb  --model=unknownq-link0 --id 1804
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=dream-panda-orb  --model=unknownq-link4 --id 1804
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=dream-panda-orb  --model=unknownq-link9 --id 1804
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=dream-panda-orb  --model=unknownq-random_all --id 1804
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=dream-panda-orb  --model=unknownq-random_top5 --id 1804
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=dream-panda-orb  --model=unknownq-random_top3 --id 1804

# Analysis of number of iterations, Supplementary Material.
python -m robopose.scripts.download --models=ablation_train_iterations
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=dream-panda-orb  --model=train_K=1 --id 1804
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=dream-panda-orb  --model=train_K=2 --id 1804
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=dream-panda-orb  --model=train_K=3 --id 1804
runjob --ngpus=8  python scripts/run_robot_eval.py --datasets=dream-panda-orb  --model=train_K=5 --id 1804

Re-training the models

We provide all the training code.

Background images for data augmentation

We apply data augmentation to the training images. Data augmentation includes pasting random images of the pascal VOC dataset on the background of the scenes. You can download Pascal VOC using the following commands:

cd local_data
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar
tar -xvf VOCtrainval_11-May-2012.tar

(If the website is down, which happens periodically, you can alternatively download these files from a mirror at https://pjreddie.com/media/files/VOCtrainval_11-May-2012.tar)

Reproducing models from the paper

runjob --ngpus=44  python scripts/run_articulated_training.py --config=dream-panda-gt_joints
runjob --ngpus=44  python scripts/run_articulated_training.py --config=dream-panda-predict_joints

runjob --ngpus=44  python scripts/run_articulated_training.py --config=dream-baxter-gt_joints
runjob --ngpus=44  python scripts/run_articulated_training.py --config=dream-baxter-predict_joints

runjob --ngpus=44  python scripts/run_articulated_training.py --config=dream-kuka-gt_joints
runjob --ngpus=44  python scripts/run_articulated_training.py --config=dream-kuka-predict_joints

runjob --ngpus=44  python scripts/run_articulated_training.py --config=craves-owi535-predict_joints