Skip to content

Commit

Permalink
Rewrote real robot section
Browse files Browse the repository at this point in the history
  • Loading branch information
@abhaybd
abhaybd committed Apr 18, 2024
1 parent 2d83a74 commit 15e6213
Show file tree
Hide file tree
Showing 2 changed files with 83 additions and 98 deletions.
169 changes: 71 additions & 98 deletions index.html
View file
Original file line number Diff line number Diff line change
Expand Up @@ -380,11 +380,75 @@ <h4 class="title is-4">Filtering High-Error Labels</h4>
</div>
</section>

<hr>

<section class="section">
<div class="container has-text-centered is-max-desktop">
<div class="columns is-centered has-text-centered is-multiline">
<div class="column is-four-fifths">
<h3 class="title is-3">Real Robot Experiments</h3>
<div class="columns is-centered">
<div class="column is-4 has-text-centered">
<img class="no_outline" src="./static/images/cube.JPEG" alt="Grasp cube task" style="max-width: 100%; max-height: 100%;">
</div>
<div class="column is-4 has-text-centered">
<img class="no_outline" src="./static/images/lego.JPEG" alt="Lego insert task" style="max-width: 100%; max-height: 100%;">
</div>
<div class="column is-4 has-text-centered">
<img class="no_outline" src="./static/images/coin.JPEG" alt="Grasp coin task" style="max-width: 100%; max-height: 100%;">
</div>
</div>
<h4 class="title is-4">CCIL improves imitation learning, especially in low-data regimes</h4>
<div class="columns is-centered">
<div class="column is-4 has-text-centered">
<img src="./static/images/IROS_figures/grasp_cube_ablation_line_plot.png" alt="Cube data ablation" style="max-width: 100%; max-height: 100%;">
</div>
<div class="column is-4 has-text-centered">
<img src="./static/images/IROS_figures/lego_insert_ablation_line_plot.png" alt="Lego data ablation" style="max-width: 100%; max-height: 100%;">
</div>
<div class="column is-4 has-text-centered">
<img src="./static/images/IROS_figures/coin_ablation_line_plot.png" alt="Coin data ablation" style="max-width: 100%; max-height: 100%;">
</div>
</div>
<div class="columns is-centered">
<div class="column is-9 has-text-centered insight">
<p>
CCIL consistently out-performs standard Behavior Cloning. This boost is especially prominent in low-data regimes, showcasing
CCIL's data efficiency and robustness.
</p>
</div>
</div>
<h4 class="title is-4">CCIL's Robustness to Lipschitz Constraint</h4>
<div class="columns is-centered is-vcentered">
<div class="column is-4 has-text-centered">
<img src="./static/images/IROS_figures/hparam_grid.png" alt="Hyperparameter ablation grid" style="max-width: 100%; max-height: 100%;">
</div>
<div class="column is-4 has-text-centered">
<img src="./static/images/IROS_figures/local_L.png" alt="Local L" style="max-width: 100%; max-height: 100%;">
</div>
</div>
<div class="columns is-centered">
<div class="column is-9 has-text-centered insight">
<p>
CCIL is relatively robust to the Lipschitz constraint. We see that for a wide range of choices of the Lipschitz constraint,
there is a choice of label error threshold that yields a significant performance boost for CCIL.
</p>
<p>
This is corroborated by the observation that different Lipschitz constraints yield similar distributions of local Lipschitz constants,
indicating that the model is able to capture the environment's inherent continuity without strict regularization.
</p>
</div>
</div>
</div>
</div>
</div>
</section>

<hr>
<section class="section">
<div class="container has-text-centered">
<!-- Animation. -->
<h2 class="title is-3">Results</h2>
<h2 class="title is-3">Simulation Experiments</h2>
<div class="columns is-centered has-text-centered is-multiline">
<div class="column is-four-fifths">
<h3 class="title is-4">Mujoco and Metaworld</h3>
Expand Down Expand Up @@ -570,108 +634,17 @@ <h3 class="title is-4">F1Tenth</h3>
</div>
</section>

<hr>

<section class="section">
<div class="container has-text-centered is-max-desktop">
<div class="columns is-centered has-text-centered is-multiline">
<div class="column is-four-fifths">
<h3 class="title is-3">Real Robot</h3>
<div class="horizontal_list">
<span class="list_image">
<img class="no_outline" src="./static/images/coin.JPEG" alt="Grasp coin task" style="max-width: 100%; max-height: 100%;">
</span>
<span class="list_image">
<img class="no_outline" src="./static/images/cube.JPEG" alt="Grasp cube task" style="max-width: 100%; max-height: 100%;">
</span>
<span class="list_image">
<img class="no_outline" src="./static/images/lego.JPEG" alt="Lego insert task" style="max-width: 100%; max-height: 100%;">
</span>
</div>
<h4 class="subtitle is-4"><strong>H.1 </strong>CCIL improves imitation learning for real-world fine-manipulation tasks, especially in low-data scenarios.</h4>
<div class="inside-hypothesis">
<div class="horizontal_list">
<span class="list_image">
<img src="./static/images/IROS_figures/grasp_cube_ablation_line_plot.png" alt="Cube data ablation" style="max-width: 100%; max-height: 100%;">
<!-- <p class="image_caption"></p> -->
</span>
<span class="list_image">
<img src="./static/images/IROS_figures/lego_insert_ablation_line_plot.png" alt="Lego data ablation" style="max-width: 100%; max-height: 100%;">
<!-- <p class="image_caption"></p> -->
</span>
</div>
<div class="horizontal_list">
<span class="list_image">
<img src="./static/images/IROS_figures/coin_ablation_line_plot.png" alt="Coin data ablation" style="max-width: 100%; max-height: 100%;">
<!-- <p class="image_caption"></p> -->
</span>
<p class="list_image image_caption description">
Across our three main tasks, we show that CCIL consistently out-performers standard Behavior Cloning. This performance boost is made especially clear in low-data regions, like in our Cube Grasping task with about a 60% improvement in success at 20% demonstrations
</p>
</div>
</div>
<h4 class="subtitle is-4"><strong>H.2 </strong>Real-world tasks with complex dynamics maintain local Lipschitz continuity, supporting CCIL's assumptions.</h4>
<div class="inside-hypothesis">
<div class="horizontal_list">
<span class="list_image">
<img src="./static/images/IROS_figures/error_scatter.png" alt="Error scatter plot" style="max-width: 100%; max-height: 100%;">
<!-- <p class="image_caption"></p> -->
</span>
<span class="list_image">
<img src="./static/images/IROS_figures/cluster_ablation.png" alt="Local Lipchitz vs. Lipschitz Constraints" style="max-width: 100%; max-height: 100%;">
<!-- <p class="image_caption"></p> -->
</span>
</div>
<div class="horizontal_list">
<span class="list_image">
<img src="./static/images/IROS_figures/local_L.png" alt="Cluster ablation" style="max-width: 100%; max-height: 100%;">
<!-- <p class="image_caption"></p> -->
</span>
<p class="list_image image_caption description">
Tasks containing local Lipschitz continuity can be realized by learning dynamics models. This model and the generated labels have varying continuity that correlate with the real world dynamics. For example, the green cluster of labels correspond with cube grasping, and the blue represents free space motion.
</p>
</div>
</div>
<h4 class="subtitle is-4"><strong>H.3 </strong>CCIL's performance is sensitive to the label rejection hyperparameter, affecting the acceptable error bound for labels.</h4>
<div class="inside-hypothesis">
<div class="horizontal_list">
<span class="list_image">
<img src="./static/images/IROS_figures/err_quantile_line_plot.png" alt="Error quantile line plot" style="max-width: 100%; max-height: 100%;">
<!-- <p class="image_caption"></p> -->
</span>
<p class="list_image image_caption description">
Label filtration thresholds are crucial to understand and tune propertly, and a balance must be struck between label accuracy and usefulness to maximize CCIL's performance.
</p>
</div>
</div>
<h4 class="subtitle is-4"><strong>H.4 </strong>CCIL also shows relative robustness to variations in the local Lipschitz constraint during dynamics model training.</h4>
<div class="inside-hypothesis">
<div class="horizontal_list">
<span class="list_image">
<img src="./static/images/IROS_figures/hparam_grid.png" alt="Hyperparameter ablation grid" style="max-width: 100%; max-height: 100%;">
<p class="image_caption">Different Lipshitz constraints will have different ideal rejection quantiles.</p>
</span>
<p class="list_image image_caption description">
CCIL is robust to different Lipschitz constraints. We see that for each Lipschitz constraint, there is a label rejection quantile that achieves a significant boost in performance over BC. CCIL can also work well with an unconstrained dynamics model. As seen in hypothesis 2, even when a Lipschitz constraint is infinite, CCIL still achieves good performance.
</p>
</div>
</div>
</div>
</div>
</div>
</section>

<hr>
<section class="section" id="BibTeX">
<div class="container is-max-desktop content">
<h2 class="title">BibTeX</h2>
<pre><code>@misc{ke2023ccil,
title={CCIL: Continuity-based Data Augmentation for Corrective Imitation Learning},
author={Liyiming Ke and Yunchu Zhang and Abhay Deshpande and Siddhartha Srinivasa and Abhishek Gupta},
year={2023},
eprint={2310.12972},
archivePrefix={arXiv},
primaryClass={cs.RO}
title={CCIL: Continuity-based Data Augmentation for Corrective Imitation Learning},
author={Liyiming Ke and Yunchu Zhang and Abhay Deshpande and Siddhartha Srinivasa and Abhishek Gupta},
year={2023},
eprint={2310.12972},
archivePrefix={arXiv},
primaryClass={cs.RO}
}</code></pre>
</div>
</section>
Expand Down
12 changes: 12 additions & 0 deletions static/css/index.css
View file
Original file line number Diff line number Diff line change
Expand Up @@ -217,3 +217,15 @@ body {
#accordion .icon {
fill: white;
}

.insight {
background-color: peachpuff;
border: 2px solid gray;
border-radius: 5px;
padding: 5px;
margin: 0 auto;
}

.insight p:not(:last-child) {
margin-bottom: 5px;
}

0 comments on commit 15e6213

Please sign in to comment.