Add heading and wrist teleoperation checkpoint

README.md

@@ -137,7 +137,19 @@ SONIC includes a kinematic planner for real-time locomotion generation — choos
 
 ## Model Checkpoints
 
-All checkpoints (ONNX format) are available directly in this repository. Inference is powered by TensorRT and runs on both desktop and Jetson hardware.
+The repository provides three Unitree G1 controller variants. Each ONNX
+encoder must be used with the decoder and observation config from the same
+variant.
+
+| Model | Location | SMPL reference input | Intended use and comments |
+|---|---|---|---|
+| **Default SONIC (original release)** | Top-level deployment files; PyTorch checkpoint under `sonic_release/` | 10 frames at 20 ms spacing, approximately 200 ms | General-purpose motion tracking, planning, and teleoperation. G1 and teleoperation references use `step5`. |
+| **Low-latency teleoperation** | `low_latency/` | 4 frames at 20 ms spacing, approximately 80 ms | Reduced-reference-lookahead controller for responsive teleoperation and VLA execution. G1 and teleoperation references use `step1`. |
+| **Heading/wrist teleoperation** | `heading_wrist/` | 10 frames at 20 ms spacing, approximately 200 ms | Robot-heading-normalized controller trained with wrist-pose augmentation for heading-stable 3-point teleoperation and SONIC-backed VLA policies. G1 and teleoperation references use `step5`; this is not the low-latency model. |
+
+All variants use 64-dimensional universal motion tokens and a 29-DoF action
+decoder. Inference is powered by TensorRT and runs on both desktop and Jetson
+hardware.
 
 | Checkpoint | File | Description |
 |---|---|---|
@@ -146,6 +158,8 @@ All checkpoints (ONNX format) are available directly in this repository. Inferen
 | Kinematic planner | `planner_sonic.onnx` | Real-time locomotion style planner |
 | Low-latency policy | `low_latency/model_encoder.onnx`, `low_latency/model_decoder.onnx`, `low_latency/observation_config.yaml` | Reduced-lookahead G1 controller variant |
 | Low-latency PyTorch checkpoint | `low_latency/last.pt` | Training checkpoint and config for the low-latency variant |
+| Heading/wrist policy | `heading_wrist/model_encoder.onnx`, `heading_wrist/model_decoder.onnx`, `heading_wrist/observation_config.yaml` | Heading-normalized G1 controller variant |
+| Heading/wrist PyTorch checkpoint | `heading_wrist/last.pt` | Training checkpoint and configs for the heading/wrist variant |
 
 **Quick download** (requires `pip install huggingface_hub`):
 
@@ -160,10 +174,34 @@ Or use the download script from the GitHub repo:
 python download_from_hf.py             # policy + planner (default)
 python download_from_hf.py --no-planner # policy only
 python download_from_hf.py --low-latency # low-latency policy + planner
+python download_from_hf.py --heading-wrist # heading/wrist policy + planner
 ```
 
 See the [Download Models guide](https://nvlabs.github.io/GR00T-WholeBodyControl/getting_started/download_models.html) for full instructions.
 
+## Heading/Wrist Inference
+
+The `heading_wrist/` variant uses robot-heading-normalized target orientation
+and a 10-frame SMPL/wrist reference horizon. It was trained with wrist-pose
+augmentation and is intended for 3-point teleoperation and SONIC-backed VLA
+policies trained against this controller.
+
+```bash
+git clone https://github.com/NVlabs/GR00T-WholeBodyControl.git
+cd GR00T-WholeBodyControl
+pip install huggingface_hub
+python download_from_hf.py --heading-wrist
+
+cd gear_sonic_deploy
+./deploy.sh \
+    --cp policy/heading_wrist/model \
+    --obs-config policy/heading_wrist/observation_config.yaml \
+    sim
+```
+
+For the Python VLA launcher, pass the same two paths with
+`--deploy-checkpoint` and `--deploy-obs-config`.
+
 ## Low-Latency Inference
 
 The low-latency SONIC variant is stored under `low_latency/` and does not

heading_wrist/config.yaml

@@ -0,0 +1,946 @@
+num_gpus: 1
+max_retries: 1
+seed: 0
+codebase_version: 1.0
+headless: true
+num_envs: 4096
+auto_load_latest: false
+checkpoint: logs_rl/TRL_G1_Track/manager/universal_token/all_modes/sonic_release_3pt_heading_wrist_81-20260415_051436/last.pt
+project_name: TRL_G1_Track
+experiment_name: ${exp_base}_${exp_var}
+base_dir: logs_rl
+timestamp: ${now:%Y%m%d_%H%M%S}
+experiment_dir: logs_rl/TRL_G1_Track/manager/universal_token/all_modes/sonic_release_3pt_heading_wrist_81-20260415_051436
+save_dir: ${experiment_dir}/.hydra
+force_flat_terrain: false
+use_wandb: true
+log_task_name: TEST
+multi_gpu: true
+global_rank: 0
+sim_type: isaacsim
+env_spacing: 20
+output_dir: ${experiment_dir}/output
+eval_overrides:
+  headless: false
+  num_envs: 1
+  auto_load_latest: false
+  use_wandb: false
+algo:
+  trl:
+    dataset_name: dummy
+    output_dir: logs_rl/TRL_G1_Track/manager/universal_token/all_modes/sonic_release_3pt_heading_wrist_81-20260415_051436
+    learning_rate: ${algo.config.actor_learning_rate}
+    num_total_batches: ${algo.config.num_learning_iterations}
+    total_episodes: null
+    num_ppo_epochs: ${algo.config.num_learning_epochs}
+    num_mini_batches: ${algo.config.num_mini_batches}
+    per_device_train_batch_size: null
+    gradient_accumulation_steps: 1
+    report_to: none
+    gamma: ${algo.config.gamma}
+    lam: ${algo.config.lam}
+    vf_coef: ${algo.config.value_loss_coef}
+    max_grad_norm: ${algo.config.max_grad_norm}
+    lr_scheduler_type: constant
+    save_strategy: 'no'
+    disable_tqdm: true
+  config:
+    critic:
+      _target_: groot.rl.trl.modules.actor_critic_modules.Critic
+      running_mean_std: true
+      backbone:
+        _target_: groot.rl.trl.modules.base_module.BaseModule
+        process_output_dim: true
+        module_config_dict:
+          type: MLP
+          input_dim:
+          - critic_obs
+          output_dim:
+          - 1
+          layer_config:
+            type: MLP
+            hidden_dims:
+            - 4096
+            - 4096
+            - 2048
+            - 2048
+            - 1024
+            - 1024
+            - 512
+            - 512
+            activation: SiLU
+    actor:
+      backbone:
+        quantizer:
+          _target_: vector_quantize_pytorch.FSQ
+        encoders:
+          g1:
+            inputs:
+            - command_multi_future_nonflat
+            - motion_anchor_ori_heading_mf_nonflat
+            params:
+              _target_: groot.rl.trl.modules.base_module.BaseModule
+              input_dim: null
+              output_dim: null
+              num_input_temporal_dims: ${manager_env.commands.motion.num_future_frames}
+              num_output_temporal_dims: ${algo.config.actor.backbone.max_num_tokens}
+              module_config_dict:
+                layer_config:
+                  type: MLP
+                  hidden_dims:
+                  - 2048
+                  - 1024
+                  - 512
+                  - 512
+                  activation: SiLU
+          teleop:
+            inputs:
+            - command_multi_future_lower_body
+            - vr_3point_local_target
+            - vr_3point_local_orn_target
+            - motion_anchor_ori_heading
+            params:
+              _target_: groot.rl.trl.modules.base_module.BaseModule
+              input_dim: null
+              output_dim: null
+              num_input_temporal_dims: null
+              num_output_temporal_dims: ${algo.config.actor.backbone.max_num_tokens}
+              module_config_dict:
+                layer_config:
+                  type: MLP
+                  hidden_dims:
+                  - 2048
+                  - 1024
+                  - 512
+                  - 512
+                  activation: SiLU
+          smpl:
+            inputs:
+            - smpl_joints_multi_future_local_nonflat
+            - smpl_root_ori_heading_multi_future
+            - joint_pos_multi_future_wrist_for_smpl
+            params:
+              _target_: groot.rl.trl.modules.base_module.BaseModule
+              input_dim: null
+              output_dim: null
+              num_input_temporal_dims: ${manager_env.commands.motion.smpl_num_future_frames}
+              num_output_temporal_dims: ${algo.config.actor.backbone.max_num_tokens}
+              module_config_dict:
+                layer_config:
+                  type: MLP
+                  hidden_dims:
+                  - 2048
+                  - 1024
+                  - 512
+                  - 512
+                  activation: SiLU
+        decoders:
+          g1_dyn:
+            inputs:
+            - token_flattened
+            - proprioception
+            outputs:
+            - action
+            has_temporal_dim: false
+            params:
+              _target_: groot.rl.trl.modules.base_module.BaseModule
+              input_dim: null
+              output_dim: null
+              num_input_temporal_dims: null
+              num_output_temporal_dims: null
+              module_config_dict:
+                layer_config:
+                  type: MLP
+                  hidden_dims:
+                  - 4096
+                  - 4096
+                  - 2048
+                  - 2048
+                  - 1024
+                  - 1024
+                  - 512
+                  - 512
+                  activation: SiLU
+          g1_kin:
+            inputs:
+            - token
+            outputs:
+            - command_multi_future_nonflat
+            - motion_anchor_ori_heading_mf_nonflat
+            has_temporal_dim: true
+            params:
+              _target_: groot.rl.trl.modules.base_module.BaseModule
+              input_dim: null
+              output_dim: null
+              num_input_temporal_dims: ${algo.config.actor.backbone.max_num_tokens}
+              num_output_temporal_dims: ${manager_env.commands.motion.num_future_frames}
+              module_config_dict:
+                layer_config:
+                  type: MLP
+                  hidden_dims:
+                  - 2048
+                  - 1024
+                  - 512
+                  - 512
+                  activation: SiLU
+        _target_: groot.rl.trl.modules.universal_token_modules.UniversalTokenModule
+        num_future_frames: ${manager_env.commands.motion.num_future_frames}
+        proprioception_features:
+        - actor_obs
+        encoder_sample_probs: ${manager_env.commands.motion.encoder_sample_probs}
+        num_fsq_levels: 32
+        fsq_level_list: 32
+        max_num_tokens: 2
+        aux_loss_func:
+          g1_recon:
+            _target_: groot.rl.trl.losses.token_losses.G1ReconLoss
+            loss_type: mse
+          g1_smpl_latent:
+            _target_: groot.rl.trl.losses.token_losses.G1SmplLatentLoss
+            loss_type: mse
+          g1_teleop_latent:
+            _target_: groot.rl.trl.losses.token_losses.G1TeleopLatentLoss
+            loss_type: mse
+          teleop_smpl_latent:
+            _target_: groot.rl.trl.losses.token_losses.TeleopSmplLatentLoss
+            loss_type: mse
+            detach_teleop_target: true
+          reencoded_smpl_g1_latent:
+            _target_: groot.rl.trl.losses.token_losses.ReencodedSmplG1LatentLoss
+            loss_type: mse
+        aux_loss_coef:
+          g1_recon: 0.01
+          g1_smpl_latent: 1.0
+          g1_teleop_latent: 1.0
+          teleop_smpl_latent: 1.0
+          reencoded_smpl_g1_latent: 1.0
+        reencode_smpl_g1_recon: true
+      _target_: groot.rl.trl.modules.actor_critic_modules.Actor
+      running_mean_std: false
+      input_obs_dict: true
+      has_aux_loss: true
+    num_learning_epochs: 5
+    num_mini_batches: 4
+    clip_param: 0.2
+    gamma: 0.99
+    lam: 0.95
+    value_loss_coef: 1.0
+    entropy_coef: 0.01
+    actor_learning_rate: 2.0e-05
+    critic_learning_rate: 0.001
+    max_grad_norm: 0.1
+    use_clipped_value_loss: true
+    schedule: adaptive
+    desired_kl: 0.01
+    adaptive_lr_min: 1.0e-05
+    adaptive_lr_max: 0.0002
+    use_new_actor_critic: true
+    use_padding_mask: false
+    ppo_shuffle_every_epoch: true
+    sync_advantage_normalization: true
+    empty_cache_every_n_ppo_epoch: -1
+    num_steps_per_env: 24
+    save_interval: 500
+    eval_frequency: 500
+    load_optimizer: true
+    init_noise_std: 0.05
+    num_learning_iterations: 100000
+    init_at_random_ep_len: true
+    eval_callbacks:
+      analysis_plot_cb:
+        _target_: groot.rl.agents.callbacks.analysis_plot_motion_tracking.AnalysisPlotMotionTracking
+        config:
+          sim_dt: 0.02
+          log_single_robot: false
+          plot_update_interval: 100
+          compute_metrics: true
+          use_weblogger: false
+    global_rank: 0
+    world_size: 128
+    use_clampped_std: true
+    std_clamp_min: 0.001
+    std_clamp_max: 0.5
+  _target_: groot.rl.agents.ppo.ppo_im.PPOIM
+  _recursive_: false
+env: ???
+terrain: ???
+callbacks:
+  model_save:
+    _target_: groot.rl.trl.callbacks.model_save_callback.ModelSaveCallback
+    save_dir: ${experiment_dir}
+    save_frequency: 2000
+    max_disk_usage: 14.8
+  wandb:
+    _target_: groot.rl.trl.callbacks.wandb_callback.WandbCallback
+  read_eval:
+    _target_: groot.rl.trl.callbacks.read_eval_callback.ReadEvalCallback
+    eval_dir: ${experiment_dir}/eval
+    check_interval: 1
+  im_resample:
+    _target_: groot.rl.trl.callbacks.im_resample_callback.ImResampleCallback
+    motion_resample_frequency: 250
+trainer:
+  _target_: groot.rl.trl.trainer.ppo_trainer_aux_loss.TRLAuxLossPPOTrainer
+wandb:
+  wandb_project: null
+  wandb_tags: online
+  wandb_group: zen_heading_0223
+  wandb_id: jbpdpb7r
+  wandb_entity: null
+  wandb_dir: /tmp
+manager_env:
+  commands:
+    motion:
+      encoder_sample_probs:
+        g1: 1.0
+        teleop: 1.0
+        smpl: 1.0
+      _target_: groot.rl.envs.manager_env.mdp.commands.TrackingCommandCfg
+      asset_name: robot
+      motion_lib_cfg:
+        motion_file:
+        - ${shared_dir}/wbc_data/training_mixes/train_pyroki_v081_0414_26/
+        smpl_motion_file:
+        - ${shared_dir}/wbc_data/smpl_retarget/1215/
+        asset:
+          assetRoot: groot/rl/data/assets/bm/unitree_description/mjcf/
+          assetFileName: g1_29dof_rev_1_0.xml
+          urdfFileName: ''
+        extend_config: []
+        target_fps: 50
+        multi_thread: true
+        filter_motion_keys: null
+        adaptive_sampling:
+          enable: true
+          bin_size: 50
+          sequence_length_agnostic: true
+          init_num_failures: 1
+          uniform_sampling_rate: 0.1
+          pre_failure_sample_window: 200
+          use_failure_rate_decay: false
+          decay_gamma: 0.8
+          adp_samp_failure_rate_max_over_mean: 200
+        upper_body_augment_prefixes:
+        - '2025'
+        - sonic_mixed_walking_running
+        - sonic_squating_different_height_switching
+        - sonic_running
+        - sonic_walking
+        - sonic_squating_different_height_hold
+        - sonic_walking_other_style
+        - balance
+        smpl_y_up: true
+      resampling_time_range:
+      - 1000000000.0
+      - 1000000000.0
+      debug_vis: true
+      dt_future_ref_frames: 0.1
+      num_future_frames: 10
+      pose_range:
+        x:
+        - -0.05
+        - 0.05
+        'y':
+        - -0.05
+        - 0.05
+        z:
+        - -0.01
+        - 0.01
+        roll:
+        - -0.1
+        - 0.1
+        pitch:
+        - -0.1
+        - 0.1
+        yaw:
+        - -0.2
+        - 0.2
+      velocity_range:
+        x:
+        - -0.5
+        - 0.5
+        'y':
+        - -0.5
+        - 0.5
+        z:
+        - -0.2
+        - 0.2
+        roll:
+        - -0.52
+        - 0.52
+        pitch:
+        - -0.52
+        - 0.52
+        yaw:
+        - -0.78
+        - 0.78
+      joint_position_range:
+      - -0.1
+      - 0.1
+      joint_velocity_range:
+      - 0
+      - 0
+      anchor_body: pelvis
+      vr_3point_body:
+      - left_wrist_yaw_link
+      - right_wrist_yaw_link
+      - torso_link
+      vr_3point_body_offset:
+      - - 0.18
+        - -0.025
+        - 0.0
+      - - 0.18
+        - 0.025
+        - 0.0
+      - - 0.0
+        - 0.0
+        - 0.35
+      reward_point_body:
+      - torso_link
+      - left_wrist_yaw_link
+      - right_wrist_yaw_link
+      reward_point_body_offset:
+      - - 0.0
+        - 0.0
+        - 0.5
+      - - 0.0
+        - -0.0
+        - 0.0
+      - - 0.0
+        - -0.0
+        - 0.0
+      body_names:
+      - pelvis
+      - left_hip_roll_link
+      - left_knee_link
+      - left_ankle_roll_link
+      - right_hip_roll_link
+      - right_knee_link
+      - right_ankle_roll_link
+      - torso_link
+      - left_shoulder_roll_link
+      - left_elbow_link
+      - left_wrist_yaw_link
+      - right_shoulder_roll_link
+      - right_elbow_link
+      - right_wrist_yaw_link
+      smpl_num_future_frames: 10
+      smpl_dt_future_ref_frames: 0.02
+      cat_upper_body_poses: true
+      cat_upper_body_poses_prob: 0.5
+      freeze_frame_aug: true
+      teleop_sample_prob_when_smpl: 0.5
+      randomize_wrist_poses: true
+    _target_: groot.rl.envs.manager_env.mdp.commands.CommandsCfg
+  actions:
+    joint_pos:
+      _target_: isaaclab.envs.mdp.actions.JointPositionActionCfg
+      asset_name: robot
+      joint_names:
+      - .*
+      use_default_offset: true
+    _target_: groot.rl.envs.manager_env.mdp.actions.ActionsCfg
+  observations:
+    policy:
+      gravity_dir:
+        _target_: isaaclab.managers.ObservationTermCfg
+        func: groot.rl.envs.manager_env.mdp:gravity_dir
+        params:
+          command_name: motion
+        history_length: ${actor_prop_history_length}
+        noise:
+          _target_: isaaclab.utils.noise.AdditiveUniformNoiseCfg
+          n_min: -0.05
+          n_max: 0.05
+      base_ang_vel:
+        _target_: isaaclab.managers.ObservationTermCfg
+        func: groot.rl.envs.manager_env.mdp:base_ang_vel
+        history_length: ${actor_prop_history_length}
+        noise:
+          _target_: isaaclab.utils.noise.AdditiveUniformNoiseCfg
+          n_min: -0.2
+          n_max: 0.2
+      joint_pos:
+        _target_: isaaclab.managers.ObservationTermCfg
+        func: groot.rl.envs.manager_env.mdp:joint_pos_rel
+        history_length: ${actor_prop_history_length}
+        noise:
+          _target_: isaaclab.utils.noise.AdditiveUniformNoiseCfg
+          n_min: -0.01
+          n_max: 0.01
+      joint_vel:
+        _target_: isaaclab.managers.ObservationTermCfg
+        func: groot.rl.envs.manager_env.mdp:joint_vel_rel
+        history_length: ${actor_prop_history_length}
+        noise:
+          _target_: isaaclab.utils.noise.AdditiveUniformNoiseCfg
+          n_min: -0.5
+          n_max: 0.5
+      actions:
+        _target_: isaaclab.managers.ObservationTermCfg
+        func: groot.rl.envs.manager_env.mdp:last_action
+        history_length: ${actor_actions_history_length}
+      _target_: groot.rl.envs.manager_env.mdp.observations.PolicyCfg
+      enable_corruption: true
+      concatenate_terms: true
+    critic:
+      command_multi_future:
+        _target_: isaaclab.managers.ObservationTermCfg
+        func: groot.rl.envs.manager_env.mdp:command_multi_future
+        params:
+          command_name: motion
+      motion_anchor_pos_b:
+        _target_: isaaclab.managers.ObservationTermCfg
+        func: groot.rl.envs.manager_env.mdp:motion_anchor_pos_b
+        params:
+          command_name: motion
+      motion_anchor_ori_b:
+        _target_: isaaclab.managers.ObservationTermCfg
+        func: groot.rl.envs.manager_env.mdp:motion_anchor_ori_b
+        params:
+          command_name: motion
+      body_pos:
+        _target_: isaaclab.managers.ObservationTermCfg
+        func: groot.rl.envs.manager_env.mdp:robot_body_pos_b
+        params:
+          command_name: motion
+      body_ori:
+        _target_: isaaclab.managers.ObservationTermCfg
+        func: groot.rl.envs.manager_env.mdp:robot_body_ori_b
+        params:
+          command_name: motion
+      base_lin_vel:
+        _target_: isaaclab.managers.ObservationTermCfg
+        func: groot.rl.envs.manager_env.mdp:base_lin_vel
+        history_length: ${critic_prop_history_length}
+      base_ang_vel:
+        _target_: isaaclab.managers.ObservationTermCfg
+        func: groot.rl.envs.manager_env.mdp:base_ang_vel
+        history_length: ${critic_prop_history_length}
+      joint_pos:
+        _target_: isaaclab.managers.ObservationTermCfg
+        func: groot.rl.envs.manager_env.mdp:joint_pos_rel
+        history_length: ${critic_prop_history_length}
+      joint_vel:
+        _target_: isaaclab.managers.ObservationTermCfg
+        func: groot.rl.envs.manager_env.mdp:joint_vel_rel
+        history_length: ${critic_prop_history_length}
+      actions:
+        _target_: isaaclab.managers.ObservationTermCfg
+        func: groot.rl.envs.manager_env.mdp:last_action
+        history_length: ${critic_actions_history_length}
+      _target_: groot.rl.envs.manager_env.mdp.observations.PrivilegedCfg
+    tokenizer:
+      encoder_index:
+        _target_: isaaclab.managers.ObservationTermCfg
+        func: groot.rl.envs.manager_env.mdp:get_command_obs
+        params:
+          command_name: motion
+          obs_name: episode_encoder_index
+      command_multi_future_nonflat:
+        _target_: isaaclab.managers.ObservationTermCfg
+        func: groot.rl.envs.manager_env.mdp:command_multi_future
+        params:
+          command_name: motion
+          non_flatten: true
+      command_z_multi_future_nonflat:
+        _target_: isaaclab.managers.ObservationTermCfg
+        func: groot.rl.envs.manager_env.mdp:command_z_multi_future
+        params:
+          command_name: motion
+          non_flatten: true
+      motion_anchor_ori_heading_mf_nonflat:
+        _target_: isaaclab.managers.ObservationTermCfg
+        func: groot.rl.envs.manager_env.mdp:motion_anchor_ori_heading_mf
+        params:
+          command_name: motion
+          non_flatten: true
+        noise:
+          _target_: isaaclab.utils.noise.AdditiveUniformNoiseCfg
+          n_min: -0.05
+          n_max: 0.05
+      command_multi_future_lower_body:
+        _target_: isaaclab.managers.ObservationTermCfg
+        func: groot.rl.envs.manager_env.mdp:command_multi_future_lower_body
+        params:
+          command_name: motion
+      vr_3point_local_target:
+        _target_: isaaclab.managers.ObservationTermCfg
+        func: groot.rl.envs.manager_env.mdp:vr_3point_local_target
+        params:
+          command_name: motion
+      vr_3point_local_orn_target:
+        _target_: isaaclab.managers.ObservationTermCfg
+        func: groot.rl.envs.manager_env.mdp:vr_3point_local_orn_target
+        params:
+          command_name: motion
+      motion_anchor_ori_heading:
+        _target_: isaaclab.managers.ObservationTermCfg
+        func: groot.rl.envs.manager_env.mdp:motion_anchor_ori_heading
+        params:
+          command_name: motion
+        noise:
+          _target_: isaaclab.utils.noise.AdditiveUniformNoiseCfg
+          n_min: -0.05
+          n_max: 0.05
+      command_z:
+        _target_: isaaclab.managers.ObservationTermCfg
+        func: groot.rl.envs.manager_env.mdp:command_z
+        params:
+          command_name: motion
+      smpl_joints_multi_future_local_nonflat:
+        _target_: isaaclab.managers.ObservationTermCfg
+        func: groot.rl.envs.manager_env.mdp:smpl_joints_multi_future_local
+        params:
+          command_name: motion
+          non_flatten: true
+        noise:
+          _target_: isaaclab.utils.noise.AdditiveUniformNoiseCfg
+          n_min: -0.05
+          n_max: 0.05
+      smpl_root_ori_heading_multi_future:
+        _target_: isaaclab.managers.ObservationTermCfg
+        func: groot.rl.envs.manager_env.mdp:smpl_root_ori_heading_mf
+        params:
+          command_name: motion
+          non_flatten: true
+        noise:
+          _target_: isaaclab.utils.noise.AdditiveUniformNoiseCfg
+          n_min: -0.05
+          n_max: 0.05
+      joint_pos_multi_future_wrist_for_smpl:
+        _target_: isaaclab.managers.ObservationTermCfg
+        func: groot.rl.envs.manager_env.mdp:joint_pos_multi_future_select_joints_for_smpl
+        params:
+          command_name: motion
+          joints_idx:
+          - 23
+          - 24
+          - 25
+          - 26
+          - 27
+          - 28
+      _target_: groot.rl.envs.manager_env.mdp.observations.TokenizerCfg
+      enable_corruption: true
+      concatenate_terms: false
+    _target_: groot.rl.envs.manager_env.mdp.observations.ObservationsCfg
+  rewards:
+    tracking_anchor_pos:
+      _target_: isaaclab.managers.RewardTermCfg
+      func: groot.rl.envs.manager_env.mdp:tracking_anchor_pos_error
+      weight: 0.5
+      params:
+        command_name: motion
+        std: 0.3
+    tracking_anchor_ori:
+      _target_: isaaclab.managers.RewardTermCfg
+      func: groot.rl.envs.manager_env.mdp:tracking_anchor_ori_error
+      weight: 0.5
+      params:
+        command_name: motion
+        std: 0.4
+    tracking_relative_body_pos:
+      _target_: isaaclab.managers.RewardTermCfg
+      func: groot.rl.envs.manager_env.mdp:tracking_relative_body_pos_error
+      weight: 1.0
+      params:
+        command_name: motion
+        std: 0.3
+    tracking_relative_body_ori:
+      _target_: isaaclab.managers.RewardTermCfg
+      func: groot.rl.envs.manager_env.mdp:tracking_relative_body_ori_error
+      weight: 1.0
+      params:
+        command_name: motion
+        std: 0.4
+    tracking_body_linvel:
+      _target_: isaaclab.managers.RewardTermCfg
+      func: groot.rl.envs.manager_env.mdp:tracking_body_linvel_error
+      weight: 1.0
+      params:
+        command_name: motion
+        std: 1.0
+    tracking_body_angvel:
+      _target_: isaaclab.managers.RewardTermCfg
+      func: groot.rl.envs.manager_env.mdp:tracking_body_angvel_error
+      weight: 1.0
+      params:
+        command_name: motion
+        std: 3.14
+    action_rate_l2:
+      _target_: isaaclab.managers.RewardTermCfg
+      func: groot.rl.envs.manager_env.mdp:action_rate_l2
+      weight: -0.1
+    joint_limit:
+      _target_: isaaclab.managers.RewardTermCfg
+      func: groot.rl.envs.manager_env.mdp:joint_pos_limits
+      weight: -10.0
+      params:
+        asset_cfg:
+          _target_: isaaclab.managers.SceneEntityCfg
+          name: robot
+          joint_names:
+          - .*
+    undesired_contacts:
+      _target_: isaaclab.managers.RewardTermCfg
+      func: groot.rl.envs.manager_env.mdp:undesired_contacts
+      weight: -0.1
+      params:
+        sensor_cfg:
+          _target_: isaaclab.managers.SceneEntityCfg
+          name: contact_forces
+          body_names:
+          - ^(?!left_ankle_roll_link$)(?!right_ankle_roll_link$)(?!left_wrist_yaw_link$)(?!right_wrist_yaw_link$)(?!left_elbow_link$)(?!right_elbow_link$).+$
+        threshold: 1.0
+    anti_shake_ang_vel:
+      _target_: isaaclab.managers.RewardTermCfg
+      func: groot.rl.envs.manager_env.mdp:anti_shake_ang_vel_l2
+      weight: -0.005
+      params:
+        command_name: motion
+        threshold: 1.5
+        body_names:
+        - left_wrist_yaw_link
+        - right_wrist_yaw_link
+        - head_link
+    tracking_vr_5point_local:
+      _target_: isaaclab.managers.RewardTermCfg
+      func: groot.rl.envs.manager_env.mdp:tracking_local_vr_5point_error
+      weight: 2.0
+      params:
+        command_name: motion
+        std: 0.1
+    tracking_vr_2wrists_local_ori:
+      _target_: isaaclab.managers.RewardTermCfg
+      func: groot.rl.envs.manager_env.mdp:tracking_local_vr_2wrists_ori_error
+      weight: 0.4
+      params:
+        command_name: motion
+        std: 0.1
+        body_names:
+        - left_wrist_yaw_link
+        - right_wrist_yaw_link
+    feet_acc:
+      _target_: isaaclab.managers.RewardTermCfg
+      func: groot.rl.envs.manager_env.mdp:joint_acc_l2
+      weight: -2.5e-06
+      params:
+        asset_cfg:
+          _target_: isaaclab.managers.SceneEntityCfg
+          name: robot
+          joint_names:
+          - .*ankle.*
+    energy_consumption:
+      _target_: isaaclab.managers.RewardTermCfg
+      func: groot.rl.envs.manager_env.mdp:energy_consumption
+      weight: -0.0001
+      params:
+        asset_cfg:
+          name: robot
+    _target_: groot.rl.envs.manager_env.mdp.rewards.RewardsCfg
+  terminations:
+    anchor_pos:
+      _target_: isaaclab.managers.TerminationTermCfg
+      func: groot.rl.envs.manager_env.mdp:exceeded_anchor_height
+      params:
+        command_name: motion
+        threshold: 0.15
+        threshold_adaptive: true
+        down_threshold: 0.75
+        root_height_threshold: 0.5
+    anchor_ori_full:
+      _target_: isaaclab.managers.TerminationTermCfg
+      func: groot.rl.envs.manager_env.mdp:exceeded_anchor_ori
+      params:
+        asset_cfg:
+          _target_: isaaclab.managers.SceneEntityCfg
+          name: robot
+        command_name: motion
+        threshold: 0.2
+    ee_body_pos:
+      _target_: isaaclab.managers.TerminationTermCfg
+      func: groot.rl.envs.manager_env.mdp:exceeded_body_height
+      params:
+        command_name: motion
+        threshold: 0.15
+        threshold_adaptive: true
+        down_threshold: 0.75
+        root_height_threshold: 0.5
+        body_names:
+        - left_ankle_roll_link
+        - right_ankle_roll_link
+        - left_wrist_yaw_link
+        - right_wrist_yaw_link
+    time_out:
+      _target_: isaaclab.managers.TerminationTermCfg
+      func: groot.rl.envs.manager_env.mdp:tracking_time_out
+      time_out: true
+      params:
+        command_name: motion
+    foot_pos_xyz:
+      _target_: isaaclab.managers.TerminationTermCfg
+      func: groot.rl.envs.manager_env.mdp:exceeded_body_pos
+      params:
+        command_name: motion
+        threshold: 0.2
+        body_names:
+        - left_ankle_roll_link
+        - right_ankle_roll_link
+    _target_: groot.rl.envs.manager_env.mdp.terminations.TerminationsCfg
+  events:
+    physics_material:
+      _target_: isaaclab.managers.EventTermCfg
+      func: groot.rl.envs.manager_env.mdp:randomize_rigid_body_material
+      mode: startup
+      params:
+        asset_cfg:
+          _target_: isaaclab.managers.SceneEntityCfg
+          name: robot
+          body_names: .*
+        static_friction_range:
+        - 0.3
+        - 1.6
+        dynamic_friction_range:
+        - 0.3
+        - 1.2
+        restitution_range:
+        - 0.0
+        - 0.5
+        num_buckets: 64
+    add_joint_default_pos:
+      _target_: isaaclab.managers.EventTermCfg
+      func: groot.rl.envs.manager_env.mdp:randomize_joint_default_pos
+      mode: startup
+      params:
+        asset_cfg:
+          _target_: isaaclab.managers.SceneEntityCfg
+          name: robot
+          joint_names:
+          - .*
+        pos_distribution_params:
+        - -0.01
+        - 0.01
+        operation: add
+    base_com:
+      _target_: isaaclab.managers.EventTermCfg
+      func: groot.rl.envs.manager_env.mdp:randomize_rigid_body_com
+      mode: startup
+      params:
+        asset_cfg:
+          _target_: isaaclab.managers.SceneEntityCfg
+          name: robot
+          body_names: torso_link
+        com_range:
+          x:
+          - -0.025
+          - 0.025
+          'y':
+          - -0.05
+          - 0.05
+          z:
+          - -0.05
+          - 0.05
+    push_robot:
+      _target_: isaaclab.managers.EventTermCfg
+      func: groot.rl.envs.manager_env.mdp:push_by_setting_velocity
+      mode: interval
+      interval_range_s:
+      - 4.0
+      - 6.0
+      params:
+        velocity_range:
+          x:
+          - -0.5
+          - 0.5
+          'y':
+          - -0.5
+          - 0.5
+          z:
+          - -0.2
+          - 0.2
+          roll:
+          - -0.52
+          - 0.52
+          pitch:
+          - -0.52
+          - 0.52
+          yaw:
+          - -0.78
+          - 0.78
+    randomize_rigid_body_mass:
+      _target_: isaaclab.managers.EventTermCfg
+      func: groot.rl.envs.manager_env.mdp:randomize_rigid_body_mass
+      mode: startup
+      params:
+        asset_cfg:
+          _target_: isaaclab.managers.SceneEntityCfg
+          name: robot
+          body_names: .*wrist_yaw.*|torso_link
+        mass_distribution_params:
+        - 0.8
+        - 2.5
+        operation: scale
+    _target_: groot.rl.envs.manager_env.mdp.events.EventCfg
+  curriculum:
+    _target_: groot.rl.envs.manager_env.mdp.curriculum.CurriculumCfg
+  recorders:
+    _target_: groot.rl.envs.manager_env.mdp.recorders.RecordersCfg
+  _target_: groot.rl.envs.manager_env.modular_tracking_env_cfg.ModularTrackingEnvCfg
+  config:
+    num_envs: ${num_envs}
+    env_spacing: 2.0
+    decimation: 4
+    episode_length_s: 10.0
+    sim_dt: 0.005
+    viewer_eye:
+    - 4.5
+    - 0.0
+    - 6.0
+    viewer_lookat:
+    - 0.0
+    - 0.0
+    - 2.0
+    render_results: false
+    save_rendering_dir: ${experiment_dir}/renderings_training
+    experiment_dir: ${experiment_dir}
+    terrain_type: trimesh
+    object_mass: null
+    table_offset: null
+    eval_camera_offset:
+    - 2
+    - 2
+    - 1
+    action_transform_module_cfg: null
+    action_transform_module_checkpoint: null
+    meta_action_dim: null
+    needs_policy_atm: true
+    obs:
+      obs_dict: {}
+      obs_dims: {}
+      group_obs_dims: {}
+      group_obs_names: {}
+    action_clip_value: 20.0
+    robot:
+      type: g1_model_12_dex
+      algo_obs_dim_dict: {}
+    rewards:
+      num_critics: 1
+    train_only_events:
+    - push_robot
+    - compliance_force_push
+use_manager_env: true
+exp_base: ${hydra:runtime.choices.exp}
+exp_var: wrist_81
+actor_prop_history_length: 10
+actor_actions_history_length: 10
+critic_prop_history_length: 10
+critic_actions_history_length: 10
+eval_datasets:
+- ${shared_dir}/wbc_data/pyroki/pyroki_v0.8.1_splits/v0.8.1_test_content
+- ${shared_dir}/wbc_data/pyroki/pyroki_v0.8.1_splits/v0.8.1_test_repetition
+eval_modes:
+- g1
+- smpl
+- teleop
+shared_dir: /mnt/amlfs-04/shared
+resume: true

heading_wrist/last.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:af24831ae59424a0cf92cb56e9bb6dc1a59ab859fd055ba13187e9e6f0a59f43
+size 1138314059

heading_wrist/model_config.yaml

@@ -0,0 +1,328 @@
+algo_config:
+  actor:
+    _target_: groot.rl.trl.modules.actor_critic_modules.Actor
+    backbone:
+      _target_: groot.rl.trl.modules.universal_token_modules.UniversalTokenModule
+      aux_loss_coef:
+        g1_recon: 0.01
+        g1_smpl_latent: 1.0
+        g1_teleop_latent: 1.0
+        reencoded_smpl_g1_latent: 1.0
+        teleop_smpl_latent: 1.0
+      aux_loss_func:
+        g1_recon:
+          _target_: groot.rl.trl.losses.token_losses.G1ReconLoss
+          loss_type: mse
+        g1_smpl_latent:
+          _target_: groot.rl.trl.losses.token_losses.G1SmplLatentLoss
+          loss_type: mse
+        g1_teleop_latent:
+          _target_: groot.rl.trl.losses.token_losses.G1TeleopLatentLoss
+          loss_type: mse
+        reencoded_smpl_g1_latent:
+          _target_: groot.rl.trl.losses.token_losses.ReencodedSmplG1LatentLoss
+          loss_type: mse
+        teleop_smpl_latent:
+          _target_: groot.rl.trl.losses.token_losses.TeleopSmplLatentLoss
+          detach_teleop_target: true
+          loss_type: mse
+      decoders:
+        g1_dyn:
+          has_temporal_dim: false
+          inputs:
+          - token_flattened
+          - proprioception
+          outputs:
+          - action
+          params:
+            _target_: groot.rl.trl.modules.base_module.BaseModule
+            input_dim: null
+            module_config_dict:
+              layer_config:
+                activation: SiLU
+                hidden_dims:
+                - 4096
+                - 4096
+                - 2048
+                - 2048
+                - 1024
+                - 1024
+                - 512
+                - 512
+                type: MLP
+            num_input_temporal_dims: null
+            num_output_temporal_dims: null
+            output_dim: null
+        g1_kin:
+          has_temporal_dim: true
+          inputs:
+          - token
+          outputs:
+          - command_multi_future_nonflat
+          - motion_anchor_ori_heading_mf_nonflat
+          params:
+            _target_: groot.rl.trl.modules.base_module.BaseModule
+            input_dim: null
+            module_config_dict:
+              layer_config:
+                activation: SiLU
+                hidden_dims:
+                - 2048
+                - 1024
+                - 512
+                - 512
+                type: MLP
+            num_input_temporal_dims: 2
+            num_output_temporal_dims: 10
+            output_dim: null
+      encoder_sample_probs:
+        g1: 1.0
+        smpl: 1.0
+        teleop: 1.0
+      encoders:
+        g1:
+          inputs:
+          - command_multi_future_nonflat
+          - motion_anchor_ori_heading_mf_nonflat
+          params:
+            _target_: groot.rl.trl.modules.base_module.BaseModule
+            input_dim: null
+            module_config_dict:
+              layer_config:
+                activation: SiLU
+                hidden_dims:
+                - 2048
+                - 1024
+                - 512
+                - 512
+                type: MLP
+            num_input_temporal_dims: 10
+            num_output_temporal_dims: 2
+            output_dim: null
+        smpl:
+          inputs:
+          - smpl_joints_multi_future_local_nonflat
+          - smpl_root_ori_heading_multi_future
+          - joint_pos_multi_future_wrist_for_smpl
+          params:
+            _target_: groot.rl.trl.modules.base_module.BaseModule
+            input_dim: null
+            module_config_dict:
+              layer_config:
+                activation: SiLU
+                hidden_dims:
+                - 2048
+                - 1024
+                - 512
+                - 512
+                type: MLP
+            num_input_temporal_dims: 10
+            num_output_temporal_dims: 2
+            output_dim: null
+        teleop:
+          inputs:
+          - command_multi_future_lower_body
+          - vr_3point_local_target
+          - vr_3point_local_orn_target
+          - motion_anchor_ori_heading
+          params:
+            _target_: groot.rl.trl.modules.base_module.BaseModule
+            input_dim: null
+            module_config_dict:
+              layer_config:
+                activation: SiLU
+                hidden_dims:
+                - 2048
+                - 1024
+                - 512
+                - 512
+                type: MLP
+            num_input_temporal_dims: null
+            num_output_temporal_dims: 2
+            output_dim: null
+      fsq_level_list: 32
+      max_num_tokens: 2
+      num_fsq_levels: 32
+      num_future_frames: 10
+      proprioception_features:
+      - actor_obs
+      quantizer:
+        _target_: vector_quantize_pytorch.FSQ
+      reencode_smpl_g1_recon: true
+    has_aux_loss: true
+    input_obs_dict: true
+    running_mean_std: false
+  actor_learning_rate: 2.0e-05
+  adaptive_lr_max: 0.0002
+  adaptive_lr_min: 1.0e-05
+  clip_param: 0.2
+  critic:
+    _target_: groot.rl.trl.modules.actor_critic_modules.Critic
+    backbone:
+      _target_: groot.rl.trl.modules.base_module.BaseModule
+      module_config_dict:
+        input_dim:
+        - critic_obs
+        layer_config:
+          activation: SiLU
+          hidden_dims:
+          - 4096
+          - 4096
+          - 2048
+          - 2048
+          - 1024
+          - 1024
+          - 512
+          - 512
+          type: MLP
+        output_dim:
+        - 1
+        type: MLP
+      process_output_dim: true
+    running_mean_std: true
+  critic_learning_rate: 0.001
+  desired_kl: 0.01
+  empty_cache_every_n_ppo_epoch: -1
+  entropy_coef: 0.01
+  eval:
+    num_eval_episodes: 150
+    num_save_episodes: 200
+    save_goal_reached_only: true
+    save_trajectories: false
+    save_videos: false
+    video_save_prob: 1.0
+  eval_callbacks:
+    analysis_plot_cb:
+      _target_: groot.rl.agents.callbacks.analysis_plot_motion_tracking.AnalysisPlotMotionTracking
+      config:
+        compute_metrics: true
+        log_single_robot: false
+        plot_update_interval: 100
+        sim_dt: 0.02
+        use_weblogger: false
+  eval_frequency: 500
+  gamma: 0.99
+  global_rank: 0
+  init_at_random_ep_len: true
+  init_noise_std: 0.05
+  lam: 0.95
+  load_optimizer: true
+  max_grad_norm: 0.1
+  num_learning_epochs: 5
+  num_learning_iterations: 100000
+  num_mini_batches: 4
+  num_steps_per_env: 24
+  ppo_shuffle_every_epoch: true
+  save_interval: 500
+  schedule: adaptive
+  std_clamp_max: 0.5
+  std_clamp_min: 0.001
+  sync_advantage_normalization: true
+  use_clampped_std: true
+  use_clipped_value_loss: true
+  use_new_actor_critic: true
+  use_padding_mask: false
+  value_loss_coef: 1.0
+  world_size: 128
+env_config:
+  action_clip_value: 20.0
+  action_transform_module_cfg: null
+  action_transform_module_checkpoint: null
+  decimation: 4
+  env_spacing: 2.0
+  episode_length_s: 10.0
+  eval_camera_offset:
+  - 2
+  - 2
+  - 1
+  experiment_dir: !!python/object/apply:pathlib.PosixPath
+  - models
+  - '0430'
+  - logs_rl
+  - TRL_G1_Track
+  - manager
+  - universal_token
+  - all_modes
+  - sonic_release_3pt_heading_wrist_81-20260415_051436
+  headless: true
+  meta_action_dim: null
+  needs_policy_atm: true
+  num_envs: 1
+  object_mass: null
+  obs:
+    group_obs_dims:
+      tokenizer:
+        command_multi_future_lower_body:
+        - 240
+        command_multi_future_nonflat:
+        - 10
+        - 58
+        command_z:
+        - 1
+        command_z_multi_future_nonflat:
+        - 10
+        - 1
+        encoder_index:
+        - 3
+        joint_pos_multi_future_wrist_for_smpl:
+        - 10
+        - 6
+        motion_anchor_ori_heading:
+        - 6
+        motion_anchor_ori_heading_mf_nonflat:
+        - 10
+        - 6
+        smpl_joints_multi_future_local_nonflat:
+        - 10
+        - 72
+        smpl_root_ori_heading_multi_future:
+        - 10
+        - 6
+        vr_3point_local_orn_target:
+        - 12
+        vr_3point_local_target:
+        - 9
+    group_obs_names:
+      tokenizer:
+      - encoder_index
+      - command_multi_future_nonflat
+      - command_z_multi_future_nonflat
+      - command_z
+      - motion_anchor_ori_heading_mf_nonflat
+      - motion_anchor_ori_heading
+      - command_multi_future_lower_body
+      - vr_3point_local_target
+      - vr_3point_local_orn_target
+      - smpl_joints_multi_future_local_nonflat
+      - smpl_root_ori_heading_multi_future
+      - joint_pos_multi_future_wrist_for_smpl
+    obs_dict: {}
+    obs_dims:
+      actor_obs: 930
+      critic_obs: 1645
+      tokenizer: 1761
+  render_results: false
+  rewards:
+    num_critics: 1
+  robot:
+    actions_dim: 29
+    algo_obs_dim_dict:
+      actor_obs: 930
+      critic_obs: 1645
+      tokenizer: 1761
+    type: g1_model_12_dex
+  save_rendering_dir: models/0430/logs_rl/TRL_G1_Track/manager/universal_token/all_modes/sonic_release_3pt_heading_wrist_81-20260415_051436/renderings_training
+  sim_dt: 0.005
+  table_offset: null
+  terrain_type: trimesh
+  train_only_events:
+  - push_robot
+  - compliance_force_push
+  viewer_eye:
+  - 4.5
+  - 0.0
+  - 6.0
+  viewer_lookat:
+  - 0.0
+  - 0.0
+  - 2.0

heading_wrist/model_decoder.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:34bae8570d4a4421a5391a5c2befd745d4a02d182ec539e5f9da44c091c67509
+size 149778638

heading_wrist/model_encoder.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fb97de22819b2057b41459802128d91723d91a25f0ad73e7bfc41a9cf8365bae
+size 50117796

heading_wrist/observation_config.yaml

@@ -0,0 +1,74 @@
+# SONIC Heading/Wrist - Observation Configuration
+# =================================================
+# Matches heading_wrist/model_encoder.onnx and model_decoder.onnx.
+#
+# Encoder input: 1751   Decoder input: 994   Token: 64
+# G1/teleop references use 10 frames at step5. SMPL and wrist references use
+# 10 frames at step1. Root orientations are normalized by robot heading.
+
+observations:
+  - name: "token_state"
+    enabled: true
+  - name: "his_base_angular_velocity_10frame_step1"
+    enabled: true
+  - name: "his_body_joint_positions_10frame_step1"
+    enabled: true
+  - name: "his_body_joint_velocities_10frame_step1"
+    enabled: true
+  - name: "his_last_actions_10frame_step1"
+    enabled: true
+  - name: "his_gravity_dir_10frame_step1"
+    enabled: true
+
+encoder:
+  dimension: 64
+  use_fp16: false
+  encoder_observations:
+    - name: "encoder_mode_4"
+      enabled: true
+    - name: "motion_joint_positions_10frame_step5"
+      enabled: true
+    - name: "motion_joint_velocities_10frame_step5"
+      enabled: true
+    - name: "motion_anchor_orientation_heading_10frame_step5"
+      enabled: true
+    - name: "motion_anchor_orientation_heading"
+      enabled: true
+    - name: "motion_joint_positions_lowerbody_10frame_step5"
+      enabled: true
+    - name: "motion_joint_velocities_lowerbody_10frame_step5"
+      enabled: true
+    - name: "vr_3point_local_target"
+      enabled: true
+    - name: "vr_3point_local_orn_target"
+      enabled: true
+    - name: "smpl_joints_10frame_step1"
+      enabled: true
+    - name: "smpl_anchor_orientation_heading_10frame_step1"
+      enabled: true
+    - name: "motion_joint_positions_wrists_10frame_step1"
+      enabled: true
+  encoder_modes:
+    - name: "g1"
+      mode_id: 0
+      required_observations:
+        - encoder_mode_4
+        - motion_joint_positions_10frame_step5
+        - motion_joint_velocities_10frame_step5
+        - motion_anchor_orientation_heading_10frame_step5
+    - name: "teleop"
+      mode_id: 1
+      required_observations:
+        - encoder_mode_4
+        - motion_joint_positions_lowerbody_10frame_step5
+        - motion_joint_velocities_lowerbody_10frame_step5
+        - vr_3point_local_target
+        - vr_3point_local_orn_target
+        - motion_anchor_orientation_heading
+    - name: "smpl"
+      mode_id: 2
+      required_observations:
+        - encoder_mode_4
+        - smpl_joints_10frame_step1
+        - smpl_anchor_orientation_heading_10frame_step1
+        - motion_joint_positions_wrists_10frame_step1