Skip to content

task_params.md

Latest commit

 

History

History
152 lines (129 loc) · 5.52 KB

task_params.md

File metadata and controls

152 lines (129 loc) · 5.52 KB
Raw

Task Parameters

Task presets live under source/whole_body_tracking/whole_body_tracking/tasks/tracking/yaml/. They are the public configuration surface for the RoboNaldo curriculum.

Loading Rules

  • --yaml right_kick/task_params_3.yaml resolves relative to the tracking yaml/ directory.
  • Absolute YAML paths are also supported.
  • inherits can be used by presets to inherit another preset and override a subset of fields.
  • Every preset is schema-checked at load time. Unknown or missing top-level keys fail fast.
  • Training archives the effective preset as params/task_params.yaml, and play/eval can reuse that archived file from a local checkpoint or WandB run.

Curriculum Fields

Field Meaning
stage Either tracking or task; selects stage-aware reward values and command behavior.
motion_weight Global multiplier for imitation rewards.
goal_weight Global multiplier for soccer task rewards.
reg_weight Global multiplier for regularization terms.
start_time_sampling_fraction Limits adaptive motion-clock sampling toward the beginning of the clip.
critic_frame_index Contact/kick reference frame used by jump triggering and adaptive sampling.

Reset and Domain Randomization

Field Meaning
init_pos_range Half-width of the ball XY spawn range around ball_init_state_pos.
init_vel_range Half-width of sampled ball linear velocity.
init_pos_z_bound Upper bound for ball spawn height offset.
init_vel_z_range Vertical velocity range; normally zero for soccer.
init_yaw_range Robot yaw perturbation at reset.
static_ball_probability Probability of zeroing sampled ball velocity.
velocity_range Random push velocity ranges for robot robustness.
rigid_num PhysX GPU rigid patch budget exponent.
use_rough_terrain Enables generated rough terrain instead of a plane.
mixed_terrain Optional mixed-terrain generator settings used when use_rough_terrain is true.

right_kick/tracking_params.yaml stays on a plane for the first tracking run. right_kick/tracking_mixed_params.yaml inherits that preset and enables mixed_terrain for robustness fine-tuning. The block controls the generated terrain bank:

use_rough_terrain: true
mixed_terrain:
  terrain_bank_rows: 4
  terrain_bank_cols: 8
  max_init_terrain_level: 0
  random_rough:
    proportion: 0.8
    noise_range: [0.0, 0.008]
  slope:
    proportion: 0.1
    slope_range: [0.0, 0.005]
  inverted_slope:
    proportion: 0.1
    slope_range: [0.0, 0.005]

Increase noise_range, slope_range, or max_init_terrain_level only after the tracking prior learns reliably on the lighter preset.

Soccer Task Fields

Field Meaning
main_foot_name Contact foot used by sensors and task rewards.
ball_init_state_pos Default ball position before randomization.
ball_init_state_vel Default ball velocity before randomization.
shot_success_threshold Target-distance threshold for successful shots.
shot_valid_steps Steps after contact during which a shot can score.
goal_reward_burst_steps Duration of the transient success reward.
goal_reset_delay_steps Delay before in-episode ball reset after a success.

Stage 3 Command Fields

Field Meaning
adapt_motion_flag Adapts the anchor command toward the ball or stabilization pose.
jump_flag Jumps the motion clock near the kick frame when an incoming ball is predicted.
critical_frame_adaptive_sampling Samples kick-entry frames around critic_frame_index.
critical_frame_sampling_window Frame window used by critical-frame sampling.
anchor_xy_offset Body-frame offset for the jump trigger's controllable-area test.
threshold_r Incoming-ball trigger radius.
threshold_t Incoming-ball trigger horizon in seconds.
lambda_factor Scale for adapted anchor displacement toward the ball.
use_ontime_ball_reset Allows in-episode ball reset after a successful shot.

Overrides

reward_overrides is grouped by category:

reward_overrides:
  motion:
    motion_global_anchor_pos:
      weight_scale: 0.5
      params:
        error_threshold: 0.0
  goal:
    error_ball_to_target:
      weight_scale: 20.0
      params:
        std: 0.5

weight_scale is multiplied by the category group weight. weight sets an absolute reward weight. params updates the target reward function parameters. Reward-term-specific values, including imitation error_threshold, should stay here instead of being added as top-level shortcut fields:

reward_overrides:
  regularization:
    ee_body_pos_termination_penalty:
      params:
        warmup_threshold: 0.35
  goal:
    goal_reward_burst:
      weight: 500.0

Legacy flat aliases such as *_weight, *_std, and warmup_ee_body_pos_threshold are intentionally unsupported. The schema fails fast on those keys so presets do not grow multiple names for the same behavior.

termination_overrides is a single-level mapping from termination term name to the override for the current preset:

termination_overrides:
  anchor_pos:
    threshold: 0.25
  ee_body_pos:
    threshold:
      left_ankle_roll_link: 0.25
      right_ankle_roll_link: 0.25
      left_wrist_yaw_link: 0.25
      right_wrist_yaw_link: 0.25
  self_collision: false

Do not nest default, tracking, or task sections under termination_overrides; each YAML preset should spell out the termination settings it wants to use.