Update README.md
Browse files
README.md
CHANGED
|
@@ -10,21 +10,19 @@ tags:
|
|
| 10 |
library_name: pytorch
|
| 11 |
---
|
| 12 |
|
| 13 |
-
# SDPC Diffusion Policy
|
| 14 |
|
| 15 |
Trained checkpoint for **SDPC (Safe Diffusion Policy with Constraint)**, an image-(and goal-)conditioned
|
| 16 |
trajectory diffusion policy for quadrotor obstacle avoidance, with hard constraints enforced at
|
| 17 |
inference time via an SLSQP projection step. Developed as part of a master's thesis at Paderborn
|
| 18 |
University.
|
| 19 |
|
| 20 |
-
Code: [ashiqlathief/
|
| 21 |
|
| 22 |
## What this checkpoint is
|
| 23 |
|
| 24 |
-
A single trained run
|
| 25 |
-
|
| 26 |
-
conditioned). It predicts an 8-step action-chunk trajectory (`horizon=8`) via a 20-step DDPM
|
| 27 |
-
denoising process (`n_diffusion_steps=20`), conditioned on 1 past FPV observation (`n_obs_steps=1`)
|
| 28 |
and the relative goal vector.
|
| 29 |
|
| 30 |
The policy was trained on FPV-camera demonstrations of a quadrotor navigating a corridor around
|
|
@@ -83,10 +81,8 @@ works") for that mechanism.
|
|
| 83 |
|
| 84 |
## How to use
|
| 85 |
|
| 86 |
-
This checkpoint is meant to be loaded with the training/eval code in
|
| 87 |
-
|
| 88 |
-
`model_config.pkl`/`diffusion_config.pkl` reference model classes defined in that repo's
|
| 89 |
-
`diffuser/models/` package.
|
| 90 |
|
| 91 |
```bash
|
| 92 |
# from the dpcc-thesis1 repo root, with env_isaaclab active
|
|
@@ -102,20 +98,4 @@ import diffuser.utils as utils
|
|
| 102 |
RUN_DIR = "isaac/logs/avoiding-crazyflie/diffusion/H8_K20_Dmodels.ImagePoseCondUNet1DTemporalCondModel_Evitp_L384/7"
|
| 103 |
diff_exp = utils.load_diffusion(RUN_DIR, epoch="best", device="cuda:0")
|
| 104 |
diffusion = diff_exp.diffusion.eval()
|
| 105 |
-
```
|
| 106 |
-
|
| 107 |
-
## Limitations
|
| 108 |
-
|
| 109 |
-
- Trained and evaluated entirely in simulation on a fixed, small set of static obstacle layouts —
|
| 110 |
-
not validated for arbitrary real-world scenes.
|
| 111 |
-
- No dynamic-obstacle-avoidance guarantee: obstacle positions used at inference are either
|
| 112 |
-
ground-truth (simulation) or perceived via a depth-camera pipeline with known accuracy limits
|
| 113 |
-
(see the repo's `depth_obstacle_estimator.py` and its documented bug history).
|
| 114 |
-
- Quantitative success-rate / constraint-violation results across projection variants are reported
|
| 115 |
-
in the accompanying thesis, not restated here to avoid duplicating numbers that may be revised
|
| 116 |
-
before submission.
|
| 117 |
-
|
| 118 |
-
## Citation
|
| 119 |
-
|
| 120 |
-
If you use this checkpoint, please cite the accompanying master's thesis (Paderborn University,
|
| 121 |
-
2026) — citation details to be added on publication.
|
|
|
|
| 10 |
library_name: pytorch
|
| 11 |
---
|
| 12 |
|
| 13 |
+
# SDPC Diffusion Policy for Quadrotor Corridor Navigation
|
| 14 |
|
| 15 |
Trained checkpoint for **SDPC (Safe Diffusion Policy with Constraint)**, an image-(and goal-)conditioned
|
| 16 |
trajectory diffusion policy for quadrotor obstacle avoidance, with hard constraints enforced at
|
| 17 |
inference time via an SLSQP projection step. Developed as part of a master's thesis at Paderborn
|
| 18 |
University.
|
| 19 |
|
| 20 |
+
Code: [ashiqlathief/SDPC-imagepolicy](https://github.com/ashiqlathief/SDPC-imagepolicy)
|
| 21 |
|
| 22 |
## What this checkpoint is
|
| 23 |
|
| 24 |
+
A single trained run `H8_K20_Dmodels.ImagePoseCondUNet1DTemporalCondModel_Evitp_L384`, seed `7` i.e. the `ImagePoseCondUNet1DTemporalCondModel` denoiser (UNet + ViT-P image encoder, goal/pose
|
| 25 |
+
conditioned). It predicts an 8-step action-chunk trajectory (`horizon=8`) via a 20-step DDPM denoising process (`n_diffusion_steps=20`), conditioned on 1 past FPV observation (`n_obs_steps=1`)
|
|
|
|
|
|
|
| 26 |
and the relative goal vector.
|
| 27 |
|
| 28 |
The policy was trained on FPV-camera demonstrations of a quadrotor navigating a corridor around
|
|
|
|
| 81 |
|
| 82 |
## How to use
|
| 83 |
|
| 84 |
+
This checkpoint is meant to be loaded with the training/eval code in [ashiqlathief/SDPC-imagepolicy](https://github.com/ashiqlathief/SDPC-imagepolicy), not standalone and
|
| 85 |
+
`model_config.pkl`/`diffusion_config.pkl` reference model classes defined in that repo's `diffuser/models/` package.
|
|
|
|
|
|
|
| 86 |
|
| 87 |
```bash
|
| 88 |
# from the dpcc-thesis1 repo root, with env_isaaclab active
|
|
|
|
| 98 |
RUN_DIR = "isaac/logs/avoiding-crazyflie/diffusion/H8_K20_Dmodels.ImagePoseCondUNet1DTemporalCondModel_Evitp_L384/7"
|
| 99 |
diff_exp = utils.load_diffusion(RUN_DIR, epoch="best", device="cuda:0")
|
| 100 |
diffusion = diff_exp.diffusion.eval()
|
| 101 |
+
```
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|