FoldPath (built from MaskPlanner codebase)

This repository is a runnable implementation of FoldPath ("End-to-End Object-Centric Motion Generation via Modulated Implicit Paths") built by reusing the local MaskPlanner project files you provided.

FoldPath differs from MaskPlanner in one key way: instead of predicting unordered discrete waypoints / segments and relying on post-processing, it learns each path as a continuous function of a scalar parameter, enabling ordered, smooth path sampling directly (no concatenation stage required). This is the central shift described in the FoldPath paper: representing each path as a neural field conditioned on object features and per-path embeddings.

1) Environment setup

Clone the Repository

git clone https://github.com/CordyZZZ/fp_re.git
cd fp_re

Environment Setup

conda create -n fp python=3.10
conda activate fp
pip install torch==2.7.1+cu118 torchvision==0.22.1+cu118 torchaudio==2.7.1+cu118 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt

2) Dataset layout (PaintNet)

This code expects the same on-disk layout as MaskPlanner: In the command lines below, <DATA_ROOT> is taken as "/fileStore/windows-v2"

<DATA_ROOT>/
  train_split.json
  test_split.json
  <SAMPLE_ID_0>/
    <SAMPLE_ID_0>.obj
    trajectory.txt
  <SAMPLE_ID_1>/
    <SAMPLE_ID_1>.obj
    trajectory.txt
  ...

Notes:

• trajectory.txt should store 6D poses (position + orientation vector) in the PaintNet format.
• The FoldPath dataset wrapper splits trajectories into paths using stroke_ids (as in MaskPlanner), then linearly resamples each path to T points.

3) Training

Example (windows):

python train_foldpath.py \
  --dataset windows-v2 \
  --data_root /fileStore/windows-v2 \
  --out_dir runs/foldpath_windows_relu \
  --epochs 200 \
  --batch_size 24 \
  --lr 3e-4 \
  --activation relu

Outputs:

• runs/foldpath_windows_relu/config.json
• runs/foldpath_windows—_relu/checkpoints/last.pth

4) Generate predictions

python generate_predictions_foldpath.py \
  --checkpoint ./runs/foldpath_windows_relu/checkpoints/last.pth \
  --config ./runs/foldpath_windows_relu/config.json \
  --dataset windows-v2  \
  --data_root /fileStore/windows-v2 \
  --output_dir ./runs/foldpath_windows_relu/ \
  --split test \
  --batch_size 4

Outputs:

• runs/foldpath_windows_relu/all_predictions.npy

5) Evaluation

python eval_foldpath.py \
  --dataset windows-v2 \
  --data_root /fileStore/windows-v2 \
  --ckpt runs/foldpath_windows_relu/checkpoints/last.pth \
  --activation relu \
  --split test \
  --out_dir ./foldpath_windows_relu/ \
  --save_predictions

Outputs:

• runs/foldpath_windows_relu/metrics.json

reproduction of TABLE 1: cuboids, windows, and shelves

dataset	activation	AP_DTW	AP_DTW (paper)	AP_DTW^50	AP_DTW^50 (paper)
cuboids	relu	x	35.2	90.5	59.8
	siren	x	60.3	x	97.5
	firen	x	91.1	x	99.2
windows	relu	x	71.8		91.4
	siren	x	71.9	90	91.3
	firen	x	75.0	x	91.9
shelves	relu	x	75.4	x	88.4
	siren	x	78.0	x	89.5
	firen	x	84.3	x	91.3

reproduction of TABLE 2: containers

dataset	activation	AP_DTW^easy	AP_DTW^easy (paper)	Paint Cov.	Paint Cov. (paper)
containers	finer	x	13.7	x	91.1

PS: As noted in the paper, PCD metrics are dependent on the sampling rate and exhibit high sensitivity to outliers, rendering them unreliable and less informative in real-world scenarios. For this reason, we omit this metric from our reproduction and comparative analysis.

6) Visualization

To enhance the elegance and maintainability of this repository, I propose integrating this normalization operation into the data preprocessing pipeline.

python normalize_dataset.py \
  --data_root /fileStore/windows-v2 \
  --output_root /fileStore/windows-v2-normalized \
  --normalization per-mesh

The resulting normalized dataset is then leveraged for visualization purposes.

python render_results_foldpath.py \
  --pred_dir /workspace/tjl/foldpath_project/runs/foldpath_windows \
  --normalized_root /fileStore/windows-v2-normalized \
  --sample_dirs 1_wr1fr_1 \
  --output_dir /workspace/tjl/foldpath_project/runs/foldpath_windows/ \
  --top_k_paths 4

Outputs:

• runs/foldpath_windows/1_wr1fr_1_pred_vs_gt.png