Q1: I would like to know whether we have to use the GPUs provided (AMD W7900D) in the preliminary round or we can use our own GPUs?

Answer: The preliminary round notification contains all the information you need. Please review it carefully.

 

Q2: How can I login the testing platform equipped with AMD W7900D GPUs?

Answer: The testing platform is currently being prepared. The login information will be sent to all participants via email at a later time.

 

Q3: For the “Optimization for Numerical Relativity code AMSS-NCKU” Is it permissible to change the "GPU_Calculation" field in the input file from "no" to "yes"? Would results obtained using GPU acceleration be considered valid and accepted for evaluation?

Answer: The allowance for modifying parallelization parameters is intended as a broad permission, enabling teams to optimize runtime performance within the prescribed framework.
The GPU version has not been maintained for a long time and may contain unknown bugs. Therefore, teams opting to use the GPU version do so at their own discretion. Please note that any issues arising from the use of the GPU version will need to be resolved independently by the participating teams.
As noted, if optimization is performed, the new algorithm must be mathematically and physically equivalent to the original one. We encourage all teams to take these factors into careful consideration when configuring their systems and adjusting the code.

 

Q4: For Embodied World Model Optimization. Under the evaluation criteria overview, modifying the `run_world_model_interaction.sh` file and the `configs/inference/world_model_interaction.yaml` file is not permitted. However, the sample workflow provided by the competition runs a single case in a single scene each time. Without changing the weight configurations in the YAML file, it will not run. The original configuration sets each of the five scenes to 0.2. When running a single scene, the corresponding value should be changed to 1.0, with the others commented out. Is this modification permitted?

Answer: We advise against modifying the configuration. During inference, retain the original data_dir path （examples/world_model_interaction_prompts）from the official YAML example, which includes 5 scenarios weighted at 0.2 each. This serves solely for code verification and does not impact model inference. The actual inference scenario is defined in run_world_model_interaction.sh. Individually modifying 20 YAML files for 20 cases may impose substantial operational challenges.

 

Q5: Request for Alternative Download of UnifoLM-WMA-0 Model. We are unable to access the official Hugging Face repository to download the required model checkpoint for the Embodied World Model task. Specifically, we cannot open the link: https://huggingface.co/unitreerobotics/UnifoLM-WMA-0-Dual. Our attempts consistently result in connection errors such as "Could not resolve host" or timeouts.

Answer: Consider trying this：https://hf-mirror.com/unitreerobotics/UnifoLM-WMA-0-Dual

 

Q6: For Embodied World Model Optimization. When running the baseline of WMA on the dataset, for almost cases, the PSNR is below 25, failing to meet the correctness requirement. Is it possible that the baseline itself has some precision issues?

Answer: The PSNR accuracy in the baseline should normally be inf (infinity). We recommend that: Check whether the model was downloaded correctly, Verify if the executed shell script has been modified, Compare whether the generated video and the ground truth (GT) have the same duration, Confirm whether the one-to-one PSNR calculation script was executed properly.

 

Q7: How can I access the login information for the remote login platform?

Answer: The sponsor-provided remote login platform for the preliminary stage, equipped with AMD W7900D GPUs, is now available. The login instructions and user manual have been sent to the registered email address of the supervisor and team leader.

 

Q8: Who provides technical support for the remote login platform?

Answer: If you encounter any issues using the platform, please contact the sponsor via email: zhengwei.lou@newpower-ai.com.

 

Q9: Baseline Test workflow for Embodied World Model Optimization.

Answer: Recently, some teams have not followed the official documentation, producing various weird and unexpected results. We hope participants will carefully read the competition handbook. Here we are also providing a set of execution steps for reference. We have tested unitree_g1_pack_camera/case1 under various test environment settings.

 

Our testing process is provided below for your reference（Whether conda is used or not, the test results are consistently similar）:

1、conda create -n unifolm-wma python==3.10.18
　  conda activate unifolm-wma
　  conda install pinocchio=3.2.0 -c conda-forge -y
　  conda install ffmpeg=7.1.1 -c conda-forge
　  git clone --recurse-submodules https://github.com/unitreerobotics/unifolm-world-model-action.git
　  cd unifolm-world-model-action
　  git submodule update --init --recursive
　  pip install -e .
　  cd external/dlimp
　  pip install -e .

2、modify configs/inference/world_model_interaction.yaml
　  data_dir: 'examples/world_model_interaction_prompts'

3、sh unitree_g1_pack_camera/case1/run_world_model_interaction.sh
　  python psnr_score_for_challenge.py --gt_video
　  unitree_g1_pack_camera/case1/unitree_g1_pack_camera_case1.mp4
　  --pred_video unitree_g1_pack_camera/case1/output/inference/0_full_fs6.mp4
　  --output_file unitree_g1_pack_camera/case1/psnr_result.json

 

Q10: The committee has provided tips about optimization strategy to improve the performance and PSNR score.

Answer: Perhaps you can check the variable ‘torch.backends.cuda.matmul.allow_tf32’, and test both setting.