CALVIN

CALVIN - A benchmark for Language-Conditioned Policy Learning for Long-Horizon Robot Manipulation Tasks Oier Mees, Lukas Hermann, Erick Rosete, Wolfram Burgard

We present CALVIN (Composing Actions from Language and Vision), an open-source simulated benchmark to learn long-horizon language-conditioned tasks. Our aim is to make it possible to develop agents that can solve many robotic manipulation tasks over a long horizon, from onboard sensors, and specified only via human language. CALVIN tasks are more complex in terms of sequence length, action space, and language than existing vision-and-language task datasets and supports flexible specification of sensor suites.

💻 Quick Start

To begin, clone this repository locally

git clone --recurse-submodules https://github.com/snasiriany/calvin.git
$ export CALVIN_ROOT=$(pwd)/calvin

Install requirements:

$ cd $CALVIN_ROOT
$ conda create -n calvin_venv python=3.8  # or use virtualenv
$ conda activate calvin_venv
$ sh install.sh

If you encounter problems installing pyhash, you might have to downgrade setuptools to a version below 58.

Download dataset (choose which split you want to download with the argument D, ABC or ABCD):

$ cd $CALVIN_ROOT/dataset
$ sh download_data.sh D | ABC | ABCD

🏋️‍♂️ Train Baseline Agent

Train baseline models:

$ cd $CALVIN_ROOT/calvin_models/calvin_agent
$ python training.py datamodule.root_data_dir=/path/to/dataset/

You want to scale your training to a multi-gpu setup? Just specify the number of GPUs and DDP will automatically be used for training thanks to Pytorch Lightning. To train on all available GPUs:

$ python training.py trainer.gpus=-1

If you have access to a Slurm cluster, follow this guide.

You can use Hydra's flexible overriding system for changing hyperparameters. For example, to train a model with rgb images from both static camera and the gripper camera:

$ python training.py datamodule/observation_space=lang_rgb_static_gripper model/perceptual_encoder=gripper_cam

To train a model with RGB-D from both cameras:

$ python training.py datamodule/observation_space=lang_rgbd_both model/perceptual_encoder=RGBD_both

To train a model with rgb images from the static camera and visual tactile observations:

$ python training.py datamodule/observation_space=lang_rgb_static_tactile model/perceptual_encoder=static_RGB_tactile

To see all available hyperparameters:

$ python training.py --help

To resume a training, just override the hydra working directory :

$ python training.py hydra.run.dir=runs/my_dir

🖼️ Sensory Observations

CALVIN supports a range of sensors commonly utilized for visuomotor control:

Static camera RGB images - with shape 200x200x3.
Static camera Depth maps - with shape 200x200x1.
Gripper camera RGB images - with shape 200x200x3.
Gripper camera Depth maps - with shape 200x200x1.
Tactile image - with shape 120x160x2x3.
Proprioceptive state - EE position (3), EE orientation in euler angles (3), gripper width (1), joint positions (7), gripper action (1).

🕹️ Action Space

In CALVIN, the agent must perform closed-loop continuous control to follow unconstrained language instructions characterizing complex robot manipulation tasks, sending continuous actions to the robot at 30hz. In order to give researchers and practitioners the freedom to experiment with different action spaces, CALVIN supports the following actions spaces:

Absolute cartesian pose - EE position (3), EE orientation in euler angles (3), gripper action (1).
Relative cartesian displacement - EE position (3), EE orientation in euler angles (3), gripper action (1).
Joint action - Joint positions (7), gripper action (1).

💪 Evaluation: The Calvin Challenge

Long-horizon Multi-task Language Control (LH-MTLC)

The aim of the CALVIN benchmark is to evaluate the learning of long-horizon language-conditioned continuous control policies. In this setting, a single agent must solve complex manipulation tasks by understanding a series of unconstrained language expressions in a row, e.g., “open the drawer. . . pick up the blue block. . . now push the block into the drawer. . . now open the sliding door”. We provide an evaluation protocol with evaluation modes of varying difficulty by choosing different combinations of sensor suites and amounts of training environments. To avoid a biased initial position, the robot is reset to a neutral position before every multi-step sequence.

To evaluate a trained calvin baseline agent, run the following command:

$ cd $CALVIN_ROOT/calvin_models/calvin_agent
$ python evaluation/evaluate_policy.py --dataset_path <PATH/TO/DATASET> --train_folder <PATH/TO/TRAINING/FOLDER>

Optional arguments:

--checkpoint <PATH/TO/CHECKPOINT>: by default, the evaluation loads the last checkpoint in the training log directory. You can instead specify the path to another checkpoint by adding this to the evaluation command.
--debug: print debug information and visualize environment.

If you want to evaluate your own model architecture on the CALVIN challenge, you can implement the CustomModel class in evaluate_policy.py as an interface to your agent. You need to implement the following methods:

__init__(): gets called once at the beginning of the evaluation.
reset(): gets called at the beginning of each evaluation sequence.
step(obs, goal): gets called every step and returns the predicted action.

Then evaluate the model by running:

$ python evaluation/evaluate_policy.py --dataset_path <PATH/TO/DATASET> --custom_model

You are also free to use your own language model instead of using the precomputed language embeddings provided by CALVIN. For this, implement CustomLangEmbeddings in evaluate_policy.py and add --custom_lang_embeddings to the evaluation command.

Multi-task Language Control (MTLC)

Alternatively, you can evaluate the policy on single tasks and without resetting the robot to a neutral position. Note that this evaluation is currently only available for our baseline agent.

$ python evaluation/evaluate_policy_singlestep.py --dataset_path <PATH/TO/DATASET> --train_folder <PATH/TO/TRAINING/FOLDER> [--checkpoint <PATH/TO/CHECKPOINT>] [--debug]

Pre-trained Model

Download the MCIL model checkpoint trained on the static camera rgb images on environment D.

$ wget http://calvin.cs.uni-freiburg.de/model_weights/D_D_static_rgb_baseline.zip
$ unzip D_D_static_rgb_baseline.zip

💬 Relabeling Raw Language Annotations

You want to try learning language conditioned policies in CALVIN with a new awesome language model?

We provide an example script to relabel the annotations with different language model provided in SBert, such as the larger MPNet (paraphrase-mpnet-base-v2) or its corresponding multilingual model (paraphrase-multilingual-mpnet-base-v2). The supported options are "mini", "mpnet" and "multi". If you want to try different SBert models, just change the model name here.

cd $CALVIN_ROOT/calvin_models/calvin_agent
python utils/relabel_with_new_lang_model.py +path=$CALVIN_ROOT/dataset/task_D_D/ +name_folder=new_lang_model_folder model.nlp_model=mpnet

If you additionally want to sample different language annotations for each sequence (from the same task annotations) in the training split run the same command with the parameter reannotate=true.

📈 SOTA Models

Open-source models that outperform the MCIL baselines from CALVIN:

Contact Oier to add your model here.

Reinforcement Learning with CALVIN

Are you interested in trying reinforcement learning agents for the different manipulation tasks in the CALVIN environment? We provide a google colab to showcase how to leverage the CALVIN task indicators to learn RL agents with a sparse reward.

FAQ

Why do you use EGL rendering?

We use EGL to move the bullet rendering from cpu (which is the default) to gpu, which is much faster. This way, we can also do rollouts during the training of the agent to track its performance. By changing from cpu to gpu, the rendered textures change slightly, so be aware of this if you plan on testing pretrained models.

I am training with multiple GPUs and why am I get OOM errors during rollouts?

PyBullet only recently added an option to select which GPU to use for rendering when using EGL (fix was commited in 3c4cb80 on Oct 22, 2021, see here. If you have an old version of PyBullet, there is no way to choose the GPU, which can lead to problems on cluster nodes with multiple GPUs, because all instances would be placed on the same GPU, slowing down the rendering and potentially leading to OOM erros.

The fix introduced an environment variable EGL_VISIBLE_DEVICES (similar to CUDA_VISIBLE_DEVICES) which lets you specify the GPU device to render on. However, there is one catch: On some machines, the device ids of CUDA and EGL do not match (e.g. CUDA device 0 could be EGL device 3). We automatically handle this in our wrapper in calvin_env and find the corresponding egl device id, so you don't have to set EGL_VISIBLE_DEVICES yourself, see here.

I am not interested in the manipulation tasks recorded, can I record different demonstration with teleop?

Yes, although it is not documented right now, all the code to record data with a VR headset is present in calvin_env in https://github.com/mees/calvin_env/blob/main/calvin_env/vrdatacollector.py

Citation

If you find the dataset or code useful, please cite:

@article{calvin21,
author = {Oier Mees and Lukas Hermann and Erick Rosete-Beas and Wolfram Burgard},
title = {CALVIN: A benchmark for Language-Conditioned Policy Learning for Long-Horizon Robot Manipulation Tasks},
journal={arXiv preprint arXiv:2112.03227},
year = 2021,
}

License

MIT License

snasiriany / calvin