ashiqali98/SDPC_diffusionmodel

SDPC Diffusion Policy for Quadrotor Corridor Navigation

Trained checkpoint for SDPC (Safe Diffusion Policy with Constraint), an image-(and goal-)conditioned trajectory diffusion policy for quadrotor obstacle avoidance, with hard constraints enforced at inference time via an SLSQP projection step. Developed as part of a master's thesis at Paderborn University.

Code: ashiqlathief/SDPC-imagepolicy

What this checkpoint is

A single trained run H8_K20_Dmodels.ImagePoseCondUNet1DTemporalCondModel_Evitp_L384, seed 7 i.e. the ImagePoseCondUNet1DTemporalCondModel denoiser (UNet + ViT-P image encoder, goal/pose 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) and the relative goal vector.

The policy was trained on FPV-camera demonstrations of a quadrotor navigating a corridor around static cylinder obstacles in NVIDIA Isaac Lab, collected with a cascaded-PID controller (see the isaac/scripts/quadcopter.py data-collection script in the repo above). It is deployed either open-loop (plain diffusion sampling) or with an in-the-loop SLSQP projection step that hard-enforces obstacle avoidance and corridor bounds at every denoising step — see the repo's README ("How SDPC works") for that mechanism.

Model architecture

Component	Value
Denoiser	ImagePoseCondUNet1DTemporalCondModel (1D temporal UNet, dim=32, dim_mults=(1,2,4,8))
Image encoder	ViT-P (vit_img_size=96, vit_patch_size=8, vit_width=512, vit_depth=6, vit_heads=8)
Image conditioning dim	384
Goal/pose conditioning dim	64 (relative goal vector goal_rel)
Action horizon	8
Observation steps	1
Diffusion steps (train/inference)	20
Predicts	epsilon (noise), l2 loss
Classifier-free guidance	condition_dropout=0.25, condition_guidance_w=1.2

Training data

• Simulator: NVIDIA Isaac Lab (Isaac Sim)
• Task: navigate a straight corridor around 5 static cylindrical obstacles to a target position
• Observation: single FPV RGB image (96×96) + robot pose
• Action: relative position delta per control step
• Demonstrations: PID-controller autopilot trajectories, recorded via isaac/scripts/quadcopter.py
• Normalization: LimitsNormalizer (min/max per-dimension), embedded in this checkpoint so it can be evaluated without the original dataset present

Training procedure

Hyperparameter	Value
Batch size	8
Learning rate	1e-4 (Adam)
Gradient accumulation	2
Training steps	100,000
EMA decay	0.995
Train/test split	0.9
Loss	L2 on predicted noise, action-weighted (action_weight=10)

Files

File	Purpose
state_best.pt	Model + EMA weights at the best validation checkpoint
model_config.pkl	Denoiser architecture config (diffuser.utils.Config)
diffusion_config.pkl	GaussianDiffusion wrapper config
dataset_config.pkl	Dataset/normalizer config used at train time
trainer_config.pkl	Optimizer/training-loop config
losses.pkl	Training loss curve

How to use

This checkpoint is meant to be loaded with the training/eval code in ashiqlathief/SDPC-imagepolicy, not standalone and model_config.pkl/diffusion_config.pkl reference model classes defined in that repo's diffuser/models/ package.

# from the dpcc-thesis1 repo root, with env_isaaclab active
hf download ashiqali98/SDPC_diffusionmodel \
  --local-dir isaac/logs/avoiding-crazyflie/diffusion/H8_K20_Dmodels.ImagePoseCondUNet1DTemporalCondModel_Evitp_L384/7

python scripts/eval_craziefliepos.py   # RUN_DIR already defaults to the path above

import diffuser.utils as utils

RUN_DIR = "isaac/logs/avoiding-crazyflie/diffusion/H8_K20_Dmodels.ImagePoseCondUNet1DTemporalCondModel_Evitp_L384/7"
diff_exp = utils.load_diffusion(RUN_DIR, epoch="best", device="cuda:0")
diffusion = diff_exp.diffusion.eval()