File size: 10,542 Bytes
33c14d6 9ffd3e8 33c14d6 9ffd3e8 33c14d6 9ffd3e8 33c14d6 9ffd3e8 33c14d6 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 | # Phase 2 configuration: scripted expert, grasp validation, domain randomisation,
# LeRobot export. Every randomisation range lives here so it can be widened later.
# Units: metres, radians unless a key says _deg or _mm. [lo, hi] means uniform.
[dataset]
repo_id = "local/so101_chess_sim"
root = "data/so101_chess_sim" # relative to the project root
fps = 30 # control and recording rate. Real UVC cameras run 30 fps.
cameras = ["overhead", "rover", "wrist"]
image_width = 640 # 4:3 like the real cameras; SmolVLA pads to 512 x 512
image_height = 480
instruction = "move the piece on the red square to the blue square"
episodes = 20000 # successful episodes to collect
workers = 8
episodes_per_piece_set = 25 # a worker draws a new piece set this often
vcodec = "h264" # libsvtav1 is smaller but much slower to encode on a laptop
[overlay]
source_rgb = [230, 40, 30] # translucent red over the source square
destination_rgb = [70, 90, 255] # translucent blue over the destination square or bin
alpha = 0.5
jitter_mm = 3.0 # per camera, per episode: imitates board-calibration error
jitter_rot_deg = 2.0
jitter_scale = 0.03
wrist_frame_jitter_mm = 0.6 # extra per-frame noise on the wrist view (hand-eye error)
[piece_sets]
# Staunton sets for 25 mm squares, varied around the 46 mm-king reference meshes.
scale = [0.90, 1.08] # whole set
height_stretch = [0.95, 1.08] # taller, slimmer sets
per_kind_height = [0.96, 1.04] # proportions differ between makers
max_foot_mm = 20.0 # a foot wider than 80% of the square is rejected
[grasp]
# Fingertip height above the base, as a fraction of piece height. The jaws then
# close on the widest part just above the tips: pawn head, rook turret, bishop mitre,
# queen and king upper-body collar. Knights are pinched on the flat sides of the
# head, jaws across the direction they face (diagonal jaws let them tilt 8-16 deg).
tip_fraction = { pawn = 0.63, rook = 0.55, knight = 0.40, bishop = 0.63, queen = 0.45, king = 0.39 }
fixed_gap_mm = 0.8 # piece axis this far off the fixed finger when descending
open_margins_mm = [3.0, 2.0] # moving jaw clearance before closing; the smaller if neighbours crowd
release_margins_mm = [3.0, 1.8] # how far the jaw opens to let go, likewise
close_q = -0.17453 # "closed" command, the gripper's lower limit
[expert]
home = [0.0, -1.2, 0.95, 1.45, 0.0] # folded; clear of self-contact under jitter
home_jitter = 0.06 # per joint, start and end pose
home_gripper = [-0.05, 0.30]
speed_scale = [0.8, 1.2] # all segment durations divided by this
joint_speed = 0.8 # rad/s, free moves from and to rest
carry_speed = 0.08 # m/s; faster lets the servos lag at full stretch
vertical_speed = 0.045 # m/s, descents and lifts
travel_clearance_mm = 10.0 # carried piece base above the tallest piece nearby
# Planning clearances. Sideways moves lag behind the plan by a few mm; slow vertical
# moves (descend, grasp, lift, place, release) track within about 0.3 mm.
plan_margin_mm = 2.0 # arm to other pieces, sideways moves
carry_margin_mm = 4.0 # carried piece to other pieces, sideways moves
vertical_margin_mm = 1.0 # arm to other pieces during vertical moves
vertical_carry_margin_mm = 1.5 # carried piece, which leans a few degrees in the hand
lift_carry_margin_mm = 0.5 # carried piece during the lift, which starts beside its neighbours
place_nudge_mm = 2.5 # drop point moved this far from crowding neighbours
static_margin_mm = 1.5 # ... from the deck, camera mounts, board edge and tray
self_margin_mm = 0.5 # hand from the upper links (their hulls are already bulky)
place_drop_mm = 1.5 # piece base released this high above the board
bin_release_mm = 30.0 # piece base above the bin floor when released
retreat_mm = 45.0
plan_budget_s = 300.0 # a pick or place plan gives up after this long (successes take seconds,
# the last fallbacks near the base up to a few minutes)
[layout]
# Board and tray geometry, drawn with each piece set (every episodes_per_piece_set
# episodes). Piece sizes follow the square size, as real sets do.
square_mm = [22.0, 25.0]
border_mm = [6.0, 22.0]
board_thickness_mm = [3.0, 20.0] # vinyl mat to thick wooden board
tray_half_x_mm = [39.0, 50.0] # outer half extents
tray_half_y_mm = [30.0, 40.0]
tray_rim_mm = [12.0, 25.0]
min_slack_m = 0.003 # a board square and flush with the deck keeps this much reach to spare
[board]
# Per episode; redrawn until every square centre is within reach_m of the pan axis
# (the arm reaches 0.282 m straight down) and the board clears the deck.
robot_plays_white_probability = 0.5 # board turned half a turn: white pieces nearest the robot
yaw_deg = [-15.0, 15.0] # casually set down, not square to the robot
offset_x = [0.0, 0.05] # from flush against the deck front
offset_y = [-0.06, 0.06]
deck_gap_m = 0.002
reach_m = 0.278
[tray]
# Either side of the board, at an angle; redrawn until its drop area is within reach.
left_probability = 0.5
azimuth_deg = [30.0, 60.0] # from straight ahead, measured at the pan axis: beside the board,
radius_m = [0.18, 0.27] # not beside the deck (carries there pass the base out of reach)
yaw_deg = [-25.0, 25.0]
board_gap_m = 0.01
deck_gap_m = 0.01
rover_camera_clearance_m = 0.10 # tray centre from the rover camera housing (right side)
reach_m = 0.272
[table]
# The table slides under everything; edges and corners, with the floor in view, come
# up often (U-shaped draw).
margin_m = 0.03 # deck, board and tray at least this far inside the edge
back_edge_min_x = -0.72 # the table stays clear of the back wall
edge_bias = [0.6, 0.6] # beta distribution over the allowed range
[pieces]
offcentre_mm = 3.0 # radius of random placement error in a square
knight_yaw_deg = 25.0 # knights face the opponent within this
random_plies = [0, 70] # random legal plies from the opening
free_move_probability = 0.15 # otherwise a legal move of the side to move
capture_to_bin_probability = 0.5 # for a capture, record the bin primitive instead
pieces_in_bin = [0, 3] # captured pieces lying in the tray (three slots)
[physics]
mass_scale = [0.7, 1.4]
friction = [0.45, 1.0] # piece and jaw sliding friction
board_friction = [0.4, 0.9]
[appearance]
flat_board_probability = 0.3 # vinyl tournament board instead of wood
tint = 0.12 # multiplicative colour jitter on wood
table_specular = [0.05, 0.7] # matte to glossy laminate (bright lamp spots)
table_shininess = [0.1, 0.9]
two_tone_arm_probability = 0.15
[lighting]
# One key lamp plus up to three others. Intensities are MuJoCo diffuse values.
key_weights = { window_key = 0.3, ceiling_fill = 0.3, desk_lamp = 0.25, sun = 0.15 }
extra_probability = 0.3
max_lights = 4
max_shadows = 2
lamp_k = [2700.0, 6500.0] # warm bulb to daylight
sun_k = [5000.0, 6500.0]
ambient = [0.15, 0.5]
intensity = { window_key = [0.25, 0.8], ceiling_fill = [0.15, 0.6], ceiling_2 = [0.1, 0.5], desk_lamp = [0.3, 0.9], sun = [0.2, 0.7], fill_2 = [0.05, 0.25] }
cutoff_deg = { window_key = [50.0, 80.0], ceiling_fill = [55.0, 90.0], ceiling_2 = [55.0, 90.0], desk_lamp = [25.0, 50.0], fill_2 = [60.0, 90.0] }
ceiling_height_m = [1.0, 1.6]
window_distance_m = [0.7, 1.4]
window_height_m = [0.5, 1.3]
desk_lamp_distance_m = [0.25, 0.6]
desk_lamp_height_m = [0.25, 0.6]
sun_elevation_deg = [20.0, 70.0]
[clutter]
# Everyday things on the table, never on the board or the tray. Tall ones (laptop, mug)
# stay out of the arm's reach or behind the robot.
probability = { laptop = 0.4, phone = 0.5, mug = 0.4, notebook = 0.4, pen = 0.4 }
cable_probability = [0.7, 0.4]
cable_length_m = [0.35, 1.0]
cable_radius_m = [0.0018, 0.003]
keep_out_m = 0.02
tall_min_reach_m = 0.34
[cameras]
# Overhead: anywhere a person could mount it at home (the owner's webcam is about 65 cm
# up, looking straight down, about 44 degrees top to bottom). Azimuth 0 is the human's
# side; the robot sits at 180, which is left out.
overhead_height_m = [0.35, 0.80] # above the board
overhead_tilt_deg = [0.0, 45.0] # from straight down
overhead_azimuth_deg = [-150.0, 150.0]
overhead_roll_deg = [-15.0, 15.0]
overhead_fovy_deg = [38.0, 62.0] # common webcams, vertical
overhead_target_jitter_m = 0.03
overhead_margin = 0.02 # board and tray corners this far inside the image
# Rover and wrist cameras sit on the robot: remounting errors only.
rover_pos_jitter = 0.01
rover_rot_jitter_deg = 4.0
rover_fovy_jitter_deg = 5.0
wrist_pos_jitter = 0.002
wrist_rot_jitter_deg = 1.0
wrist_fovy_jitter_deg = 1.5
[webcam]
# Per camera, per episode: how a cheap UVC webcam renders the scene. Applied before
# the red and blue squares, which are drawn on the captured frame.
auto_exposure_target = [0.38, 0.55] # mean brightness the camera settles on
white_balance = 0.12 # per-channel gain jitter, warm to cool
exposure = [0.75, 1.3] # around the auto-exposure: under- to over-exposed
gamma = [0.85, 1.2]
saturation = [0.75, 1.2]
contrast = [0.85, 1.15]
blur_px = [0.0, 1.2] # Gaussian sigma
noise = [0.0, 0.012] # per-pixel standard deviation, per frame (room light; noise is what inflates the videos)
jpeg_quality = [45, 95]
vignette = [0.0, 0.35]
[validation]
layouts = 4 # nominal board plus random boards, each with its own piece set
piece_sets = 4 # reference set plus random sets (with their board and tray)
bin_every = 4 # every nth trial drops into the bin
success_center_mm = 6.0 # piece centre within this of the square centre
success_tilt_deg = 10.0
disturb_mm = 2.0 # no other piece moves more than this
workers = 8
|