rlmesh.adapters.Image

Semantic role matched against env image features.

Target image height, or None to keep the env height.

Target image width, or None to keep the env width.

Axis layout the model expects ("hwc" the default, or "chw"). This is the model author’s declaration: when it differs from the env’s layout the adapter transposes silently to match, so a model that wants "chw" must say so – an omitted layout means "hwc" and the env’s frame is fed through unchanged. (channels guards the channel count, not the axis order.)

Channel count the model expects (e.g. 3 for RGB, 1 for grayscale). When set, a resolve error if the env image differs; the adapter does not convert between channel counts.

NumPy dtype name the model expects.

Whether (and into what range) to map 8-bit pixel values before casting: False (off, the default), True (the conventional [0, 1]), or a (low, high) pair for a model trained on a different range (e.g. (-1.0, 1.0)). One field, so an on/off flag can never disagree with a range; False is an authoritative off-switch.

Number of leading singleton axes to add (batch/time).

Whether the model was trained on images rotated 180 degrees from the canonical upright orientation (a true rotation – rows and columns reversed – not a vertical flip). Declared on both ends: the adapter rotates only when the env and model disagree, so an env that already renders upside-down pairs with a model that also sets upside_down and no rotation happens.

Resize algorithm the model’s training pipeline used. An un-suffixed name has OpenCV/torch semantics, an _aa suffix has PIL’s (an antialiased filter whose support widens with the downscale factor): "bilinear" (4-tap half-pixel-center bilinear; the default, which most trained policies match), "bilinear_aa" / "bicubic_aa" / "lanczos3_aa" (PIL’s BILINEAR / BICUBIC / LANCZOS), or "area" (OpenCV INTER_AREA). Bare "bicubic"/"lanczos3" are not accepted – name the library whose kernel you trained against.

Permit a target larger than the env’s native resolution (interpolating detail that is not there). Off by default: an upscaling target is a resolve error unless this is set.

How to reconcile a target whose aspect ratio differs from the env image: "stretch" (distort), "crop" (cover + center-crop), or "pad" (letterbox) – or a sequence of them in preference order. The resolver picks, per env, the first that does not need a disallowed upscale, so one spec can crop a large camera and letterbox a small one. Required only on an aspect mismatch; absent it, an aspect-changing resize is a resolve error.

Zero-fill a black frame when the env does not provide this camera, instead of failing resolution. Needs height, width, and channels so the blank can be sized.

Fill value (0-255) for the blank frame an optional camera synthesizes when the env offers no matching image; default 0 (black). Requires optional=True (without it the fill could never take effect, so setting it alone is a construction error); named to match Actuator.fill.

Number of frames to stack on a new leading axis (frame history). 1 (default) means no stacking. Stacking is applied by the adapter core from an episode-keyed rolling window, padded at the start of an episode (see stack_pad) and cleared on reset. Which frames it gathers is offsets (or stride); by default they are the last stack consecutive ones. The env still sends one frame per step either way.

Side fraction of the frame a center crop keeps, in (0, 1] (2/3 keeps the middle two thirds of each axis). Pass crop or crop_area, not both.

The same center crop stated as an area fraction, in (0, 1] – the form training pipelines usually quote (“a 90% center crop”). The side fraction is its square root (0.9 -> 0.949).

How the crop box is taken. "zoom" (the default) resamples the fractional box straight to the target – PIL’s Image.resize(size, box=...), one pass, no intermediate rounding. "slice" cuts an integer center box out first and resizes that, which is what a numpy slice in a training pipeline does.

Quality of a JPEG round-trip applied to the upright frame before the crop and resize, on the IJG 1-100 scale (95 is the usual training value). A declaration, not a request: a pipeline that stored its frames as JPEG fed the model the codec’s artifacts, so the adapter reproduces them rather than handing the model a cleaner frame than it was trained on. Baseline sequential, 4:2:0 box-averaged chroma, standard IJG tables; needs a 3-channel image.

Channel order the model was trained on. "bgr" swaps red and blue after the spatial ops and before the dtype cast, and needs a 3-channel image.

The camera resolution this model was trained against, as a square int or an (height, width) pair (each axis 1-4096). An assertion, not a request: the adapter never resizes a camera to reach it. The platform binds the environment’s camera dial from it before the run; resolution then fails with RenderMismatch if the camera it bound does not actually render at this size. Declare it instead of pinning the env’s camera width/height by hand.

The frame window stack gathers, as non-positive offsets from the current step – oldest first, ending at 0. (-6, -4, -2, 0) is “every second frame of the last seven”. None (the default) is the contiguous window stack already describes. offsets and stack are never inferred from each other: both are declared and they must agree (len(offsets) == stack), and the window may span at most 128 consecutive frames.

What fills the window before the episode has produced enough frames: "first" (the default) replicates the first observed frame, "black" a raw 8-bit 0 frame pushed through this input’s pipeline – so under normalize=(-1.0, 1.0) a black pad frame is -1.0, not 0.0. Requires stack > 1.

The body part the wanted camera sits on, when the role repeats across a body ("head", "left_arm", …): an identity key the resolver matches on. Naming one binds only that camera; naming none binds the env’s only camera of the role under any part (with an info) and fails when there are several. Keyword-only and omitted from the wire when unset.