chillinfree333/stack-cups-rollouts-decoded-20260923

Decoded rollouts, 22-23 September 2026 night session

One rollout_NNNN_decoded.npz per rollout, 0276 through 0323 (48 runs, 23:00:23 to 00:58:39), plus MANIFEST.json with per-array quality statistics.

Produced by ~/yiming/tools/decode_rollouts/decode_rollout.py. Source rollouts are untouched in ~/yiming/workstation_client/rollouts/rollout_NNNN/.

What was actually decoded

The vae_* arrays in a source rollout are latents, shape (1, 48, T, 10, 20), not images. The VAE decode was already performed by the client; its output is predicted_rgb / predicted_depth / predicted_flow, each (N, 160, 320, 3) uint8. Those are still codec images: depth and flow painted into RGB. This export performs the remaining step, inverting the codecs back to metres, pixels and unit vectors, for both the predicted and the observed streams.

The codec maths is tools/teaser_recovery/codec_depth.py and codec_flow.py, verbatim ports of the cluster's own encoders, validated on rollout 0190: depth inverse median 0.18 mm, flow uint8 round trip median 0.086 px (max 0.24 px), ray round trip median 0.17 degrees.

Composite layout (verified against this data, not assumed)

Three-view runs store a DreamZero composite in every image slot:

band	rows	cols	camera
top	0-79	0-319 (one continuous tile)	wrist
bottom	80-159	0-159	left
bottom	80-159	160-319	right

Matching each tile against the separately stored per-camera frames gives wrist corr 0.9982 (MAE 0.84) and bottom-right to right corr 0.9865 (MAE 2.74), while bottom-left matches right at only corr 0.36. Column-step analysis finds a 13-27x seam at x=160 in the bottom band and no seam at x=160 in the top band, which is why the top band is one 320-wide tile and not two copies of a 160-wide one.

Left-only runs are not composites: the whole 160x320 frame is the left camera.

The flow slot is not uniform across tiles. In a three-view run the wrist tile is a world-rotation raymap (rgb = (d+1)/2 for a unit direction), while left and right carry polar-encoded optical flow. Evidence: ||2*rgb-1|| over the top band has median 1.0000 (p01 0.9951, p99 1.0049), whereas the bottom band sits at sqrt(3) with 95% pure white, the flow codec's background. Left-only runs have no raymap at all.

Arrays

Prefix pred_ for the model's output, obs_ for what the cameras saw. Camera suffix is _left, _right or _wrist (three-view) or just _left (left-only).

array	dtype	units
{pred,obs}_rgb_{cam}	uint8	the image, no transform needed
obs_rgb_{wrist,right}_full	uint8	full 160x320, higher res than the composite tile
{pred,obs}_depth_m_{cam}	float32	metres, NaN where invalid
{pred,obs}_depth_valid_{cam}	bool	exact-white test
{pred,obs}_depth_off_path_{cam}	float16	distance from the codec's RGB-cube path
obs_depth_m_zed_left	float32	the ZED's own metric depth, never codec-encoded
obs_depth_valid_zed_left	bool
{pred,obs}_flow_px_{left,right}	float16	pixels (u,v), per-view; NaN where unwritten
{pred,obs}_flow_moving_{left,right}	bool	non-white and non-black
{pred,obs}_flow_frame_observed	bool	per frame, see below
{pred,obs}_ray_unit_wrist	float16	unit world direction
{pred,obs}_ray_raw_norm_wrist	float16	norm before renormalising; 1.0 on a real raymap

Index arrays are carried through unchanged: visual_frame_action_step, visual_frame_round, visual_frame_in_round, observed_physical_action_step, observed_round, predicted_index_for_observed{,_exact}, predicted_action, predicted_action_round, executed_action, observed_state, observed_joint_positions.

Two traps this export handles for you

Two thirds of the observed flow frames were never written. They are stored as pure black. Black is not something either codec can emit (the depth codec walks the RGB cube edge path and never reaches the origin; the flow codec composites a saturated hue over a white background, so its darkest output is a pure hue), so an all-black frame means "no observation", not "zero flow". This matters because the flow decoder reads opacity as 1 - min(rgb), so black would decode to the maximum representable displacement, 53.67 px, in every pixel. Measured: 120 of 179 frames on rollout 0276, 89 of 133 on 0279, 88 of 131 on 0298, a uniform 0.67 of all pixels in every band, because flow is only built at the keyframes that feed compute_kv_cache, not at every control step. Those frames are NaN here and flagged by *_flow_frame_observed. The depth slot has no black at all; it is written every step.

A decoded depth of 4.0 m is usually an artifact. White is overloaded in the depth codec: it marks both "invalid" and "at or beyond d_max = 4 m". When the three-view composite is bilinearly downscaled, a pixel straddling a valid/invalid boundary blends toward white, lands off the codec's path, and decodes to a meaningless depth that is normally pinned at exactly 4.0 m. Filter on *_depth_off_path_* < 0.02 and the range collapses to the truth: rollout 0276's left tile goes from a nominal 4.00 m down to 2.17 m, which is exactly the ZED's own maximum of 2.166 m for the same frames. Left-only runs have no composite and no resampling, so their on-path fraction is 1.000 and their decoded depth already agrees with the ZED to the last digit (2.81 m and 2.59 m on rollouts 0279 and 0298).

Coverage is not uniform

The two RGB cells have no predicted depth and no predicted flow, because those checkpoints were trained without either. Some of them still carry observed depth and flow, because the ZED was opened in NEURAL mode regardless. The three-view RGB runs carry neither. MANIFEST.json lists the arrays each file actually contains; do not assume a key exists.