AllegroKuka.yaml

defaults:
  - _self_
  - env: reorientation

name: AllegroKuka

physics_engine: ${..physics_engine}

env:
  subtask: ""

  # if given, will override the device setting in gym.
  numEnvs: ${resolve_default:8192,${...num_envs}}

  envSpacing: 1.2
  episodeLength: 600
  enableDebugVis: False
  evalStats: False  # extra evaluation-time statistics

  clampAbsObservations: 10.0

  stiffnessScale: 1.0
  forceLimitScale: 1.0
  useRelativeControl: False
  dofSpeedScale: 10.0
  actionsMovingAverage: 1.0
  controlFrequencyInv: 1 # 60 Hz

  resetPositionNoiseX: 0.1
  resetPositionNoiseY: 0.1
  resetPositionNoiseZ: 0.02
  resetRotationNoise: 1.0
  resetDofPosRandomIntervalFingers: 0.1
  resetDofPosRandomIntervalArm: 0.1
  resetDofVelRandomInterval: 0.5

  # Random forces applied to the object
  forceScale: 2.0
  forceProbRange: [0.001, 0.1]
  forceDecay: 0.99
  forceDecayInterval: 0.08

  liftingRewScale: 20.0
  liftingBonus: 300.0
  liftingBonusThreshold: 0.15  # when the object is lifted this distance (in meters) above the table, the agent gets the lifting bonus
  keypointRewScale: 200.0
  distanceDeltaRewScale: 50.0
  reachGoalBonus: 1000.0
  kukaActionsPenaltyScale: 0.003
  allegroActionsPenaltyScale: 0.0003
  fallDistance: 0.24
  fallPenalty: 0.0

  privilegedActions: False
  privilegedActionsTorque: 0.02

  # Physics v1, pretty much default settings we used from the start of the project
  dofFriction: -1.0  # negative values are ignored and the default friction from URDF file is used

  # gain of PD controller (?)
  allegroStiffness: 40.0
  kukaStiffness: 40.0

  allegroEffort: 0.35  # this is what was used in sim-to-real experiment. Motor torque in Newton*meters
  kukaEffort: [300, 300, 300, 300, 300, 300, 300]  # see Physics v2

  allegroDamping: 5.0
  kukaDamping: 5.0

  allegroArmature: 0
  kukaArmature: 0

  keypointScale: 1.5
  objectBaseSize: 0.05

  randomizeObjectDimensions: True
  withSmallCuboids: True
  withBigCuboids: True
  withSticks: True

  objectType: "block"
  observationType: "full_state"
  successTolerance: 0.075
  targetSuccessTolerance: 0.01
  toleranceCurriculumIncrement: 0.9  # multiplicative
  toleranceCurriculumInterval: 3000  # in env steps across all agents, with 8192 this is 3000 * 8192 = 24.6M env steps
  maxConsecutiveSuccesses: 50
  successSteps: 1  # how many steps we should be within the tolerance before we declare a success

  saveStates: False
  saveStatesFile: "rootTensorsDofStates.bin"

  loadInitialStates: False
  loadStatesFile: "rootTensorsDofStates.bin"

  asset:
    # Whis was the original kuka_allegro asset.
    # This URDF has some issues, i.e. weights of fingers are too high and the mass of the Allegro hand is too
    # high in general. But in turn this leads to smoother movements and better looking behaviors.
    # Additionally, collision shapes of fingertips are more primitive (just rough convex hulls), which
    # gives a bit more FPS.
    kukaAllegro: "urdf/kuka_allegro_description/kuka_allegro_touch_sensor.urdf"

    # This is the URDF which has more accurate collision shapes and weights.
    # I believe since the hand is much lighter, the policy has more control over the movement of both arm and
    # fingers which leads to faster training (better sample efficiency). But overall the resulting
    # behaviors look too fast and a bit unrealistic.
    # For sim-to-real experiments this needs to be addressed. Overall, v2 is a "Better" URDF, and it should not
    # lead to behaviors that would be worse for sim-to-real experiments. Most likely the problem is elsewhere,
    # for example the max torques might be too high, or the armature of the motors is too low.
    # The exercise of finding the right URDF and other parameters is left for the sim-to-real part of the project.
    # kukaAllegro: "urdf/kuka_allegro_description/kuka_allegro_v2.urdf"

    assetFileNameBlock: "urdf/objects/cube_multicolor.urdf"

task:
  randomize: False
  randomization_params:
    frequency: 480   # Define how many simulation steps between generating new randomizations

    observations:
      range: [0, .002] # range for the white noise
      range_correlated: [0, .001 ] # range for correlated noise, refreshed with freq `frequency`
      operation: "additive"
      distribution: "gaussian"
      schedule: "linear"   # "constant" is to turn on noise after `schedule_steps` num steps
      schedule_steps: 40000
    actions:
      range: [0., .05]
      range_correlated: [0, .015] # range for correlated noise, refreshed with freq `frequency`
      operation: "additive"
      distribution: "gaussian"
      schedule: "linear"  # "linear" will linearly interpolate between no rand and max rand
      schedule_steps: 40000

    sim_params:
    gravity:
      range: [0, 0.4]
      operation: "additive"
      distribution: "gaussian"
      schedule: "linear"  # "linear" will linearly interpolate between no rand and max rand
      schedule_steps: 40000

    actor_params:
      allegro:
        color: True
        dof_properties:
          damping:
            range: [0.3, 3.0]
            operation: "scaling"
            distribution: "loguniform"
            schedule: "linear"  # "linear" will scale the current random sample by `min(current num steps, schedule_steps) / schedule_steps`
            schedule_steps: 30000
          stiffness:
            range: [0.75, 1.5]
            operation: "scaling"
            distribution: "loguniform"
            schedule: "linear"  # "linear" will scale the current random sample by `min(current num steps, schedule_steps) / schedule_steps`
            schedule_steps: 30000
          lower:
            range: [0, 0.01]
            operation: "additive"
            distribution: "gaussian"
            schedule: "linear"  # "linear" will scale the current random sample by `min(current num steps, schedule_steps) / schedule_steps`
            schedule_steps: 30000
          upper:
            range: [0, 0.01]
            operation: "additive"
            distribution: "gaussian"
            schedule: "linear"  # "linear" will scale the current random sample by `min(current num steps, schedule_steps) / schedule_steps`
            schedule_steps: 30000
        rigid_body_properties:
          mass:
            range: [0.5, 1.5]
            operation: "scaling"
            distribution: "uniform"
            schedule: "linear"  # "linear" will scale the current random sample by `min(current num steps, schedule_steps) / schedule_steps`
            schedule_steps: 30000
        rigid_shape_properties:
          friction:
            num_buckets: 250
            range: [0.7, 1.3]
            operation: "scaling"
            distribution: "uniform"
            schedule: "linear"  # "linear" will scale the current random sample by `min(current num steps, schedule_steps) / schedule_steps`
            schedule_steps: 30000
      object:
        scale:
          range: [0.5, 2.0]
          operation: "scaling"
          distribution: "uniform"
          schedule: "linear"  # "linear" will scale the current random sample by ``min(current num steps, schedule_steps) / schedule_steps`
          schedule_steps: 1
        rigid_body_properties:
          mass:
            range: [0.5, 1.5]
            operation: "scaling"
            distribution: "uniform"
            schedule: "linear"  # "linear" will scale the current random sample by ``min(current num steps, schedule_steps) / schedule_steps`
            schedule_steps: 30000
        rigid_shape_properties:
          friction:
            num_buckets: 250
            range: [0.7, 1.3]
            operation: "scaling"
            distribution: "uniform"
            schedule: "linear"  # "linear" will scale the current random sample by `min(current num steps, schedule_steps) / schedule_steps`
            schedule_steps: 30000

sim:
  substeps: 2
  dt: 0.01667 # 1/60
  up_axis: "z"
  use_gpu_pipeline: ${eq:${...pipeline},"gpu"}
  num_client_threads: 8

  gravity: [0.0, 0.0, -9.81]
  physx:
    num_threads: 6
    solver_type: 1  # 0: pgs, 1: tgs
    num_position_iterations: 8
    num_velocity_iterations: 0

    max_gpu_contact_pairs: 8388608 # 8*1024*1024
    num_subscenes: ${....num_subscenes}
    contact_offset: 0.002
    rest_offset: 0.0
    bounce_threshold_velocity: 0.2
    max_depenetration_velocity: 1000.0
    default_buffer_size_multiplier: 25.0
    contact_collection: 0 # 0: CC_NEVER (don't collect contact info), 1: CC_LAST_SUBSTEP (collect only contacts on last substep), 2: CC_ALL_SUBSTEPS (broken - do not use!)