File size: 3,349 Bytes
fad6045 | 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 | # Baseline profile: the simplest scene, to get reliable closed-loop success before
# adding variety back one change at a time. Merged over phase2_config.toml when
# PHASE2_PROFILE=baseline is set (only the keys listed here change).
#
# Fixed: the nominal board (square to the robot, which plays black, flush against the
# deck), reference piece set, tray on the right, table, lighting (the scene's window
# and ceiling lamps), an overhead camera like the owner's (65 cm up,
# straight down, 44 degree lens), no clutter. Black and white pieces, white arm. Outlined markers.
# Ten fixed opening moves from the start position. Sharper video (CRF 18, full
# chroma) so the markers survive encoding.
[dataset]
repo_id = "local/so101_chess_baseline"
root = "data/so101_chess_baseline"
cameras = ["overhead", "wrist"] # the two the policy uses
episodes = 300
episodes_per_piece_set = 1000 # one piece set throughout
video_crf = 18
video_pix_fmt = "yuv444p"
[overlay]
alpha = 0.6
outline_px = 2 # opaque edge around each square
jitter_mm = 0.0
jitter_rot_deg = 0.0
jitter_scale = 0.0
wrist_frame_jitter_mm = 0.0
[piece_sets]
fixed = true # reference set on the nominal board and tray
[board]
robot_plays_white_probability = 0.0
yaw_deg = [0.0, 0.0]
offset_x = [0.0, 0.0]
offset_y = [0.0, 0.0]
[tray]
left_probability = 0.0
azimuth_deg = [47.0, 47.0] # right of the board, 7 mm in from the scene file's tray
radius_m = [0.236, 0.236] # (whose far drop corner is 3 mm past the tray reach)
yaw_deg = [0.0, 0.0]
[table]
fixed_offset = [0.0, 0.0]
[pieces]
offcentre_mm = 0.0
knight_yaw_deg = 0.0
# From the start position: five for the robot's side (black, nearest the robot) and
# five for the far side. Pawns and knights, near and far, left and right.
fixed_moves = ["e7e5", "d7d5", "g8f6", "b8c6", "c7c5", "e2e4", "d2d4", "g1f3", "b1c3", "c2c4"]
[physics]
mass_scale = [1.0, 1.0]
friction = [0.7, 0.7]
board_friction = [0.6, 0.6]
[appearance]
fixed_look = true # nominal board, table, floor, walls, deck and tray colours
piece_colours = "black_white"
arm_colours = ["white"]
two_tone_arm_probability = 0.0
[lighting]
fixed = true
[clutter]
probability = { laptop = 0.0, phone = 0.0, mug = 0.0, notebook = 0.0, pen = 0.0 }
cable_probability = [0.0, 0.0]
[cameras]
# Like the owner's webcam: straight down from 65 cm over the board, 44 degrees top to
# bottom, the arm at the top edge of the image. From there the resting gripper hangs over
# the eighth rank, as it does on the real desk, so that is allowed here.
overhead_height_m = [0.65, 0.65]
overhead_tilt_deg = [0.0, 0.0]
overhead_azimuth_deg = [0.0, 0.0]
overhead_roll_deg = [0.0, 0.0]
overhead_fovy_deg = [44.0, 44.0]
overhead_target_jitter_m = 0.0
overhead_arm_may_hide = true
rover_pos_jitter = 0.0
rover_rot_jitter_deg = 0.0
rover_fovy_jitter_deg = 0.0
wrist_pos_jitter = 0.0
wrist_rot_jitter_deg = 0.0
wrist_fovy_jitter_deg = 0.0
[webcam]
# A clean camera: fixed auto-exposure target, no colour, blur or noise effects, light JPEG.
auto_exposure_target = [0.45, 0.45]
white_balance = 0.0
exposure = [1.0, 1.0]
gamma = [1.0, 1.0]
saturation = [1.0, 1.0]
contrast = [1.0, 1.0]
blur_px = [0.0, 0.0]
noise = [0.0, 0.0]
jpeg_quality = [95, 95]
vignette = [0.0, 0.0]
|