choicow/physbird-red_v2

PhysBird red_v2

Science Birds (open-source Angry Birds clone) episodes for training action-conditioned world models. Code: the physbird-wm repository (data generation, VAEs, NanoWM world models, evaluation).

• 1,584 generated levels: towers of wood / ice / stone blocks with pigs; red birds only; blocks indestructible.
• Each level is reset 12 times with a different first shot (counterfactual branches); 4 branches continue with the remaining birds. 19,008 episodes in total.
• Physics at 50 Hz; the training store keeps every 4th step (12.5 fps), frames 320 x 640 RGB, fixed camera.

split	levels	episodes	frames
train	1500	18,000	2,276,950
test (level families seen in training)	24	288	36,120
test_ood (held-out family x material cells)	60	720	89,336

Layout

red_v2_hi/                     training store (one torch file per episode)
  <split>/<hundreds:03d>/<tens:02d>/<ones>/<id>.pt
      obs_blk, obs_shape, obs_block   frames as independent zstd blocks of 32 frames, uint8 (T, 3, 320, 640)
      act      (T, 28)    action of the window after frame t: phase one-hot (no shot / aiming / launched), tap,
                          two shot slots (vx0, vy0 px, tau, pull, cos, sin, bird type one-hot 5, valid)
      state    (T, K, 9)  per body: x, y in [-1, 1], cos, sin, vx, vy px/frame, w, mass, alive
      shape    (K, 6)     w, h, kind (box / circle / triangle), material, gravity scale, drag
      regime   (T, K)     0 absent / 1 free flight / 2 contact / 3 resting / 4 scripted
      contacts (E, 6)     frame, body a, body b, relative speed, impulse, static partner
      bird     (T, 3)     pixel position of the active bird + valid
      dyn, ff, traj, activity, rew, end, trunc, info (level id, branch, split, frame indices)
  background.pt                fixed background frame, uint8 (3, 320, 640)
  files_<split>.txt            episode list per split
  manifest.json                split levels, camera, action / state layouts
red_v2/                        raw generator output (levels/*.xml, per-step episodes, stats); only needed to
                               re-convert the store
index/<vae>/nwm_index_<split>.pt   world-model episode index: per episode the store file and latent directory
                               (relative paths), length, events (launch / flight / impact / onset / collapse),
                               realigned actions, bird track
latents/<vae>/<split>/<episode>/   (optional, can be regenerated with scripts/cache_latents.sh)
  zu_d4.npy   (T, 8, 80, 160) float16   VAE posterior mean of frames (t-1, t), normalised per channel
  obj_d4.npy  (T, 64, 12) float32       GT object records: valid, class, cx, cy, long, short, cos, sin, vx, vy, omega, symmetry

<vae> is uvae_d4 (unified VAE, physics read-out on 4 of the 8 channels) or uvae_d4_vis (visual-only VAE). Read a store episode with latent.data.load_raw / decode_frames from the code repository.

Units: px, px/frame, px/frame^2. A launched bird falls at 0.96 px/frame^2.