MJX Training Implementation

.gitignore

@@ -19,3 +19,6 @@ build/
 dist/
 *.so
 out*/
+
+# Training artifacts
+wandb/

Makefile

@@ -49,8 +49,8 @@ clean:
 py-files := $(shell find . -name '*.py')
 
 format:
-	@black $(py-files)
-	@ruff format $(py-files)
+	@black ksim
+	@ruff format ksim
 .PHONY: format
 
 format-cpp:
@@ -59,13 +59,13 @@ format-cpp:
 .PHONY: format-cpp
 
 static-checks:
-	@black --diff --check $(py-files)
-	@ruff check $(py-files)
-	@mypy --install-types --non-interactive $(py-files)
+	@black --diff --check ksim
+	@ruff check ksim
+	@mypy --install-types --non-interactive ksim
 .PHONY: lint
 
 mypy-daemon:
-	@dmypy run -- $(py-files)
+	@dmypy run -- ksim
 .PHONY: mypy-daemon
 
 # ------------------------ #

README.md

@@ -21,4 +21,104 @@
 
 # K-Scale Sim
 
-A library for simulating Stompy in MuJoCo.
+A library for simulating Stompy in MJX and MuJoCo.
+
+## Installation
+1. Clone this repository:
+```bash
+git clone https://github.com/kscalelabs/ksim.git
+cd ksim
+```
+
+2. It is recommended that you use a virtual environment to install the dependencies for this project. You can create a new conda environment using the following command:
+```bash
+conda create --name ksim python=3.11
+conda activate ksim
+```
+To install the dependencies, run the following command:
+
+```bash
+make install-dev
+```
+
+## MJX Gym Usage
+MJX Gym is a library for training and evaluating reinforcement learning agents in MJX environments. It is built on top of the Brax library and provides a simple interface for running experiments with Stompy and other humanoid formfactors. Currently, we support walking but plan on adding more tasks and simulator environments in the future.
+
+
+
+### Training
+For quick experimentation, you may specify all relevant training configurations via YAML files, and simply run train.py with the desired configuration. Our configurations allow for rapid reward function prototyping, environment specification, and hyperparameter tuning. Additionally, all training results are tracked and logged using Weights & Biases.
+
+We recommend starting with the default humanoid environment to get a feel for the simulator. To train the default humanoid environment, first navigate to the /mjx_gym directory:
+```bash
+cd ksim/mjx_gym
+```
+Then, run the following command:
+```bash
+python train.py --config experiments/default_humanoid_walk.yaml
+```
+
+Example training curves are shown below:
+<p align="center">
+  <img alt="K-Scale Open Source Robotics" src="https://private-user-images.githubusercontent.com/43460304/333030841-5a4bd24e-0478-4cbb-bb92-18cf5e9c72b7.png?jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.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.Yi2HzUGt3HYivMyGTuRFSRjsZviZRK2dFEx_KtZyEyA" width="400" />
+  <img alt="K-Scale Open Source Robotics" src="https://private-user-images.githubusercontent.com/43460304/333030844-341f4a5c-95ea-49ed-b102-df4e00868c5a.png?jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.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.01D5U4QD8JooJwEDtcw9OKt1tmM_zGYj5U7dDstLC84" width="400" >
+</p>
+
+
+### Testing and Rendering
+We provide an easy way to test and render the trained model using the play.py script. This script loads the trained model and runs it in the specified environment.
+
+To test and render the trained model, run the following command:
+```bash
+python play.py --config experiments/default_humanoid_walk.yaml
+```
+
+Here is an example of the trained model walking in the default humanoid environment after roughly 15 minutes of training on the K-Scale cluster.
+
+<video controls>
+  <source alt="K-Scale Open Source Robotics" src="https://github.com/kscalelabs/sim/assets/43460304/8e12b0e6-48ea-4af0-8283-1dc4880767b4" type="video/mp4">
+</video>
+
+You may transfer the model (trained via MJX) to MuJoCo for CPU-based simulation via arguments to the play script. For example, to run the model in the default humanoid environment, use the following command:
+```bash
+python play.py --config experiments/default_humanoid_walk.yaml --use_mujoco
+```
+
+It is important that the configuration file used for testing is the **same as the one used for training.** This ensures that the model is loaded correctly and avoids any catastrophic failures during testing. For simplicity, we have included default training weights in the /mjx_gym/weights directory, which are automatically used when running the play script with the command above. However, you may specify a different weight file using the --params_path argument.
+
+Here is an example of the trained model (in MJX) walking in the default humanoid environment in MuJoCo.
+
+<video controls>
+  <source alt="K-Scale Open Source Robotics" src="https://github.com/kscalelabs/sim/assets/43460304/7f158aeb-6bc9-4056-bd1d-12882adbd13c" type="video/mp4">
+</video>
+
+While the model performs well in the MJX environment, it is important to note that the model's performance may vary when transferred to MuJoCo. This is due to slight differences in the simulators and the underlying physics engines. Evaluating the model in both simulators is helpful for determining whether the model will generalize well across different environments.
+
+### Common Issues
+You might see the following error when running train.py:
+>An NVIDIA GPU may be present on this machine, but a CUDA-enabled jaxlib is not installed. Falling back to cpu.
+
+One way to fix this issue is to uninstall jax and jaxlib and reinstall them with your specific CUDA version the following commands:
+```bash
+pip uninstall jax jaxlib
+pip install --upgrade "jax[cuda12]" -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html
+```
+Note that the CUDA version should match the one installed on your machine.
+
+A similar issue might also occur:
+> jaxlib.xla_extension.XlaRuntimeError: FAILED_PRECONDITION: DNN library initialization failed. Look at the errors above for more details.
+
+This error is usually caused by an incorrect version of cudnn installed in your environment. To fix this issue, you can try installing cuDNN using the following command:
+```bash
+conda install cudnn
+```
+
+Another common issue occurs when rendering on a headless server. For now, we recommend rendering locally or using a remote desktop connection to view the rendering.
+
+## TODO
+- [ ] Get Stompy to load efficiently with MJX (currently, the meshes and collision detection are not loading correctly)
+- [ ] Add goal conditioning to the humanoid walk task
+- [ ] Implement imitation learning techniques for end-to-end walking and recovering tasks
+- [ ] Constantly iterate on the reward functions to improve training efficiency
+- [ ] Add CPU-based MuJoCo training platform for improved sim-to-sim support
+- [ ] Add sim-to-real transfer learning techniques during training

ksim/mjx_gym/envs/__init__.py

@@ -0,0 +1,10 @@
+from brax import envs
+from envs.default_humanoid_env.default_humanoid import DefaultHumanoidEnv
+from envs.stompy_env.stompy import StompyEnv
+
+environments = {"default_humanoid": DefaultHumanoidEnv, "stompy": StompyEnv}
+
+
+def get_env(name: str, **kwargs) -> envs.Env:
+    envs.register_environment(name, environments[name])
+    return envs.get_environment(name, **kwargs)

ksim/mjx_gym/envs/default_humanoid_env/default_humanoid.py

@@ -0,0 +1,152 @@
+import jax
+import jax.numpy as jp
+import mujoco
+from brax.envs.base import PipelineEnv, State
+from brax.io import mjcf
+from brax.mjx.base import State as mjxState
+from etils import epath
+
+from envs.default_humanoid_env.rewards import DEFAULT_REWARD_PARAMS
+from envs.default_humanoid_env.rewards import get_reward_fn
+
+
+class DefaultHumanoidEnv(PipelineEnv):
+    """
+    An environment for humanoid body position, velocities, and angles.
+
+    Note: This environment is based on the default humanoid environment in the Brax library.
+    https://github.com/google/brax/blob/main/brax/envs/humanoid.py
+
+    However, this environment is designed to work with modular reward functions, allowing for quicker experimentation.
+    """
+
+    def __init__(
+        self,
+        reward_params=DEFAULT_REWARD_PARAMS,
+        terminate_when_unhealthy=True,
+        reset_noise_scale=1e-2,
+        exclude_current_positions_from_observation=True,
+        log_reward_breakdown=True,
+        **kwargs,
+    ):
+        path = epath.Path(epath.resource_path("mujoco")) / ("mjx/test_data/humanoid")
+        mj_model = mujoco.MjModel.from_xml_path((path / "humanoid.xml").as_posix())  # type: ignore
+        mj_model.opt.solver = mujoco.mjtSolver.mjSOL_CG  # type: ignore # TODO: not sure why typing is not working here
+        mj_model.opt.iterations = 6
+        mj_model.opt.ls_iterations = 6
+
+        sys = mjcf.load_model(mj_model)
+
+        physics_steps_per_control_step = 4  # Should find way to perturb this value in the future
+        kwargs["n_frames"] = kwargs.get("n_frames", physics_steps_per_control_step)
+        kwargs["backend"] = "mjx"
+
+        super().__init__(sys, **kwargs)
+
+        self._reward_params = reward_params
+        self._terminate_when_unhealthy = terminate_when_unhealthy
+        self._reset_noise_scale = reset_noise_scale
+        self._exclude_current_positions_from_observation = exclude_current_positions_from_observation
+        self._log_reward_breakdown = log_reward_breakdown
+
+        self.reward_fn = get_reward_fn(self._reward_params, self.dt, include_reward_breakdown=True)
+
+    def reset(self, rng: jp.ndarray) -> State:
+        """Resets the environment to an initial state.
+
+        Args:
+            rng: Random number generator seed.
+        Returns:
+            The initial state of the environment.
+        """
+        rng, rng1, rng2 = jax.random.split(rng, 3)
+
+        low, hi = -self._reset_noise_scale, self._reset_noise_scale
+        qpos = self.sys.qpos0 + jax.random.uniform(rng1, (self.sys.nq,), minval=low, maxval=hi)
+        qvel = jax.random.uniform(rng2, (self.sys.nv,), minval=low, maxval=hi)
+
+        mjx_state = self.pipeline_init(qpos, qvel)
+        assert type(mjx_state) == mjxState, f"mjx_state is of type {type(mjx_state)}"
+
+        obs = self._get_obs(mjx_state, jp.zeros(self.sys.nu))
+        reward, done, zero = jp.zeros(3)
+        metrics = {
+            "x_position": zero,
+            "y_position": zero,
+            "distance_from_origin": zero,
+            "x_velocity": zero,
+            "y_velocity": zero,
+        }
+        for key in self._reward_params.keys():
+            metrics[key] = zero
+
+        return State(mjx_state, obs, reward, done, metrics)
+
+    def step(self, state: State, action: jp.ndarray) -> State:
+        """Runs one timestep of the environment's dynamics.
+
+        Args:
+            state: The current state of the environment.
+            action: The action to take.
+        Returns:
+            A tuple of the next state, the reward, whether the episode has ended, and additional information.
+        """
+        mjx_state = state.pipeline_state
+        assert mjx_state, "state.pipeline_state was recorded as None"
+        # TODO: determine whether to raise an error or reset the environment
+
+        next_mjx_state = self.pipeline_step(mjx_state, action)
+
+        assert type(next_mjx_state) == mjxState, f"next_mjx_state is of type {type(next_mjx_state)}"
+        assert type(mjx_state) == mjxState, f"mjx_state is of type {type(mjx_state)}"
+        # mlutz: from what I've seen, .pipeline_state and .pipeline_step(...) actually return an brax.mjx.base.State object
+        # however, the type hinting suggests that it should return a brax.base.State object
+        # brax.mjx.base.State inherits from brax.base.State but also inherits from mjx.Data, which is needed for some rewards
+
+        obs = self._get_obs(mjx_state, action)
+        reward, is_healthy, reward_breakdown = self.reward_fn(mjx_state, action, next_mjx_state)
+
+        if self._terminate_when_unhealthy:
+            done = 1.0 - is_healthy
+        else:
+            done = jp.array(0)
+
+        state.metrics.update(
+            x_position=next_mjx_state.subtree_com[1][0],
+            y_position=next_mjx_state.subtree_com[1][1],
+            distance_from_origin=jp.linalg.norm(next_mjx_state.subtree_com[1]),
+            x_velocity=(next_mjx_state.subtree_com[1][0] - mjx_state.subtree_com[1][0]) / self.dt,
+            y_velocity=(next_mjx_state.subtree_com[1][1] - mjx_state.subtree_com[1][1]) / self.dt,
+        )
+
+        if self._log_reward_breakdown:
+            for key, val in reward_breakdown.items():
+                state.metrics[key] = val
+
+        return state.replace(  # type: ignore # TODO: fix the type hinting...
+            pipeline_state=next_mjx_state, obs=obs, reward=reward, done=done
+        )
+
+    def _get_obs(self, data: mjxState, action: jp.ndarray) -> jp.ndarray:
+        """Observes humanoid body position, velocities, and angles.
+
+        Args:
+            data: The current state of the environment.
+            action: The current action.
+        Returns:
+            Observations of the environment.
+        """
+        position = data.qpos
+        if self._exclude_current_positions_from_observation:
+            position = position[2:]
+
+        # external_contact_forces are excluded
+        return jp.concatenate(
+            [
+                position,
+                data.qvel,
+                data.cinert[1:].ravel(),
+                data.cvel[1:].ravel(),
+                data.qfrc_actuator,
+            ]
+        )