<b>Simulated log pile data</b>

Contains examples of both successful and failed grasps</p><p dir="ltr"><i>skip_config.yaml/</i><br>A yaml-file of a list of subset indices. The list contains the subset that we have manually idendified as erroneous in some way and want to exclude from the dataset generation process later on</p><p><br></p><p dir="ltr"><i>exclude_subsets.yaml/</i><br>Similar to skip_config.yaml but instead contains a list of subsets that were never considered during simulation and therefor don't have any successful grasps associated with them.

This is to tell those examples apart from the case where we tried to find good grasps but were unsuccessful.</p><h4 dir="ltr"><br><i>log_data.npz</i></h4><p dir="ltr">A dict with fields `overlap`, `contact`, `index`, `diameter`, `length`, `yaw`, `type`, `pos`, `dir` containg the all information regarding the n logs in the pile.</p><p dir="ltr">- `overlap` - An n-by-n matrix O describing which logs are obstructed by other logs.

O_ij=0 if log i and j are not obstructing each other. O_ij=1 if log i is obstructed (is under) log j. O_ij=-1 if log i is obstructing (is on top of) log j.

Logs are never obstructing themselves so O_ii is zero by definition. The matrix is necessarily anti-symmetric.</p><p dir="ltr">- `contact` - An n-by-n matrix C describing which logs are in contact with each other. O_{ij}=1 if log i and j are in contact and 0 otherwise.

Logs are never in contact with themselves so O_{ii} is zero by definition. The matrix is necessarily symmetric.</p><p dir="ltr">- `index` An array of length n with the index of each log</p><p dir="ltr">- `diameter` An array of length n with the log diameters</p><p dir="ltr">- `length` An array of length n with the log lengths</p><p dir="ltr">- `yaw` An array of length n with the logs' yaw angle relative world z-axis</p><p dir="ltr">- `type` An array of length n with the log type (str)</p><p dir="ltr">- `pos` An array of length n-by-3 with the log positions (CoM)</p><p dir="ltr">- `dir` An array of length n-by-3 with the log pose, described by a vector aligned with the log.

We can parametrize a line running along the log as</p><p dir="ltr">pos[i] + 0.5*length[i]

• dir[i]
• alpha, where alpha is in [-1, 1]. Since the logs are cylindrical we don't care about the angle around this vector.</p><p dir="ltr">image_arrays.npz</p><p dir="ltr">A dict with fields `depth`, `masks`, `rgb`, `rgb_random_0`, `rgb_random_1`, `rgb_random_2`, containg the all image data for a pile of n logs. All images in the dataset have the same extents, [-2.5, 2.5]^2 m, and the same resolution 256-by-256.</p><p dir="ltr">- `depth` A depth image of the pile. The depth data is expressed using the non-linear OpenGL scaling $$ F_\mathrm{depth}=\frac{1/z−1/\mathrm{near}}{1/\mathrm{far}−1/\mathrm{near}}$$ with $\mathrm{near}=3$ and $\mathrm{far}=6$ (learnopengl.com/Advanced-OpenGL/Depth-testing).</p><p dir="ltr">- `mask` a 256-by-256-by-(n+2) array of binary masks of the n logs as well as one mask for all tree stumps and one for all rocks.</p><p dir="ltr">- `rgb`, `rgb_old`, `rgb_random_0`, `rgb_random_1`, `rgb_random_2` various rendered images of the pile. The `_random_` images were the ones used in the paper</p><p dir="ltr"><i>heightfield.npz</i></p><p dir="ltr">A dict with fields `size` and `heights`.</p><p dir="ltr">- `size` The width and length of the terrain patch (always 10-by-10 m)</p><p dir="ltr">- `height` A 2D array of height values</p><h3 dir="ltr"> Test piles</h3><p><br></p><p dir="ltr">These are a set of 1200 piles used during quantitative testing. The test pile set is separated into groups of samples with 2, 3, ...,7 logs, and w/, w/o obstacles with a 100 piles in each group. Contains a subset of the same data found in the training piles (no ground truth grasps for example).</p><h2 dir="ltr">Grasp map dataset</h2><h3 dir="ltr">The dataset is split into two parts: `training_data/` and `validation_data/`. Both subfolders contain grasp map samples in the form of npz-files. Fields marked with a (*) are the ones used in the model from the paper.<br><br>Grasp map sample</h3><p dir="ltr"><br>A grasp map sample is stored as an npz file. The npz-file contains arrays of size $256 \times 256 \times N_c$ for some number of channels $N_c$ depending on which field we are considering. The complete list of arrays is<br>- `rgb_random` A $256 \times 256 \times 3 \times 3$ containing the rgb image of the pile rendered in three different versions. The third dimension holds the version, so rgb_random[:, :, i, :] would access the ith version.<br>- `bw_random` A $256 \times 256 \times 3$ array containing black-and-white versions of the rgb images in `rgb_random`. (*)<br>- `depth` A depth image of the pile (*)<br>- `obstacle_mask` a binary mask indicating which pixels are identified as obstacles<br>- `target_mask` a binary mask indicating which logs are targeted in this sample (*)<br>- `log_mask` a binary mask indicating which pixels are identified as logs (including targets)<br>- `indicator` a binary image showing the ground truth grasps encoded using "grasp rectangles". This image shows the grasp rectangles' centers (*)<br>- `indicator_full` same as `indicator` but shows the full span of the grasps, not just the centers.<br>- `width` encoded grasp width (*)<br>- `sin` and `cos` encoded grasp angle $\phi$. The angle is encoded using the cosine and sine of the grasp angle (*)<br>- `balance` encodes how well-balanced the resulting log bundle is inside the grapple after being grasped (*)<br>- `alignment` encodes the longitudinal alignment of the logs inside the grapple after being grasped.<br><br></p>