MimicKit

Project notes

Choose a simulator before installing the Python packages

This profile covers release v0 at commit 50b7b3d60c1ffe7f7b7159aebc273a926ce81fab, not newer unreviewed development changes. That release documents Isaac Gym and Isaac Lab backends and recommends separate Python environments for each. Install the simulator first, then requirements.txt, and set the configuration's engine_name to the matching backend. The example humanoid configuration selects Isaac Gym. [1] [2] [3] [6]

The runner defaults to cuda:0. Although its training interface accepts CPU and CUDA device names, that is not evidence that either simulator will run every workflow on a CPU. Exact simulator versions, Python versions, GPU memory and training costs need a configuration-specific check; the reviewed guide does not pin a complete working environment. [1] [4]

Try a saved controller before a large training job

The repository directs readers to a separate download for assets and motion data. The documented test uses four environments and data/models/deepmimic_humanoid_spinkick_model.pt, selected with the DeepMimic humanoid argument file. Confirm that the corresponding character, motion and model files are present before using that command. The full training example starts 4,096 environments, so it should not be treated as a minimal installation check. [1] [3]

For new motion data, inspect the clip with the view_motion environment first. Motion files store root position, root rotation and joint rotations; joint order follows the character XML's depth-first kinematic order, and three-dimensional rotations use exponential maps. A file loading successfully does not show that its joint ordering matches a different character. [1]

The example humanoid configuration uses position control at 30 Hz with simulation at 120 Hz. Those are settings for that simulated configuration, not a real-robot control contract. Training produces model parameters and logs; playback and training statistics need to be checked separately from any later hardware transfer. [1] [3]

Source and external assets have different review coverage

The repository code is Apache-2.0. We reviewed the repository source and license, but did not retrieve the separately hosted motion/asset package or pretrained model payloads. Our unauthenticated access check returned an access-denied response, which leaves a retrieval gap rather than proving general unavailability. [8] Their specific licensing and availability remain unverified; the code license is not a blanket grant for those external materials. No simulator installation, training run, model playback or robot deployment was performed. [1] [7]