Placement

BetaWorks, but under active development — APIs and results may shift.

Where things start is part of a task. A .task does not give poses; it gives distributions — “the brick starts somewhere in this region” — and every reset draws from them. Writing a fixed pose would make every episode identical, which surfaces much later as a policy that memorised one layout.

Regions

A region is declared in the .family against a site, so it moves with whatever the site is attached to:

# region=<name>:site:<site>[:xmin,ymin,xmax,ymax[:half_height]]
region=table_top:site:table_surface:-0.10,-0.10,0.10,0.10:0.03

# region=<name>:box:<site>:xmin,ymin,xmax,ymax:half_height[:<contact slot>]
region=flat_stove_1_cook_region:box:flat_stove_1_cook_region:-0.075,-0.075,0.075,0.075:0.0025:flat_stove_1

Kind	Geometry	Used for
site	An xy rectangle in the site’s frame, plus a z half-band centred on the site (default 0.12 m if omitted)	Table zones, start regions, In / On / AtRegion goals
box	The site’s own box (size read off the asset by the LIBERO importer); optional contact slot	LIBERO’s fixture regions — drawers, burners, racks — with LIBERO’s in_box / under semantics

With no rectangle, a site region is the site’s own extent. A region used by a goal must be among the first 16 the family declares (the device region table’s size); regions used only by init= are sampled on the host and have no such limit.

init= — where a free slot starts

init=brick@table_left                                    # draw inside a region
init=akita_black_bowl_1@main_table_akita_black_bowl_region
init=akita_black_bowl_1@cookies_1                        # stack on another slot

• <slot>@<region> draws a position inside the region by rejection sampling against the other placed objects.
• <slot>@<slot> stacks: the object is placed directly on the reference object at the reference’s x/y, as LIBERO’s ObjectBasedSampler does. The reference is walked first.
• :in after the target marks a placement that LIBERO spells In rather than On. The two differ by one line in LIBERO — On adds the fixture’s top_site height, In does not — and treating them the same put a bowl 4.5 cm into the stove. The generated LIBERO tasks carry :in where the .bddl says In.

A free slot the task does not list under active= is parked. The resting height and the clash radius come from the asset when the family gives a slot_geom= line (slot_geom=<slot>:bottom_z,top_z,h_radius, read from robosuite’s bottom_site, top_site and horizontal_radius_site); otherwise the caller supplies a radius.

jinit= — where a joint starts

jinit=wooden_cabinet_1_top_level@-0.16,-0.14

A uniform draw per reset in [lo, hi], for articulated fixtures: a drawer that starts part-open, a microwave door. It is a range because LIBERO’s own OpenCloseSampler is a uniform draw — a single value would make every episode start the drawer at the same opening. The range belongs to the task; the joint belongs to the family. At most two per task.

The host sampler

noeira/tasks/sampler.mojo sample_placements is the reference implementation:

• Deterministic given (task, seed, lane). Draws are counter-based (Philox seeded per draw from seed, lane, slot and attempt), so lane 7’s layout does not depend on how many attempts lanes 0–6 needed.
• Independent of the arm’s reset noise. The seed is salted away from the env’s joint-noise stream; sharing it would correlate every lane’s object layout with its starting arm pose.
• Pure geometry. It takes region frames and slot geometry as numbers and never touches Data or the model, so the same rules can run in a kernel.
• Raises on exhaustion rather than returning an overlapping scene.

noeira/tasks/reset.mojo writes the sampled poses into qpos / qvel and parks the inactive slots.

Device placement tables

The GPU reset places props per lane, inside init_qpos_gpu, with noeira/tasks/placement/table.mojo place_free_slots — the device twin of the host sampler. A kernel cannot read a .family, so it reads the family’s geometry through a comptime PlacementTable: park poses, slot addresses, slot geometry and each region’s site frame at rest.

The tables are generated for every family whose free slots all carry slot_geom= — the 23 LIBERO families and so101_tower — into <task root>/placement/:

pixi run gen-placement-tables          # write the tables
pixi run gen-placement-tables-check    # fail if any is stale (CI)

The generator uses the same inputs as the host sampler: the loaded spec, the free-slot addresses, and region frames from forward kinematics on the composed scene. Every float is written as a decimal that round-trips to the same bits, so a table can be checked for exact drift. so101_tabletop has no slot_geom= and keeps a hand-written table (So101TabletopPlacement in noeira/tasks/family_config.mojo).

Each lane’s start regions reach the kernel as one meta word per free slot (META_IDX_INIT_REGION_0 onward, up to 13), plus jinit= draws in META_IDX_JINIT_0 onward. The encoding is biased so that zero means “not placed” — the value an untouched, zero-filled meta holds — which keeps a driver that forgets to write them from silently placing every prop.

Where host and device cannot agree, the difference is visible: the host raises on exhaustion, a kernel cannot, so the kernel leaves that slot at its park pose and every goal naming it is false. Spec-level problems are refused before any word is written, by require_device_placement in noeira/tasks/placement/check.mojo.

The gate is pixi run device-placement: the device reset against the host sampler on so101_tabletop and every LIBERO task, exactly, with a count of how many of the kernel’s rules the corpus exercises.

Frozen init tables

A success rate over start states a run drew for itself is not comparable with another run’s. noeira/tasks/init_table.mojo freezes start states into a TrajectoryStore, one row per episode:

Column	Type	Content
init_state	float64 [1 + nq + nv]	time, qpos, qvel — LIBERO’s .pruned_init layout
task_index	int32	which task the episode runs
active_mask	float64	which slots that task activates

The mask is stored, not derived: an eval run months later must not change its answer because someone edited an active= line. The goal still comes from the .task, because the goal is the benchmark’s definition and belongs under version control. A table carries its family’s nq / nv in the row shape, and loading it against a different family raises.

For LIBERO, the frozen table is converted from LIBERO’s own .pruned_init files (pixi run libero-init-dump, then pixi run libero-init-freeze), and currently exists for libero_goal — see LIBERO.

Posed reset on the host

Viewers and host envs need one lane’s episode without a GPU. noeira/tasks/posed_reset.mojo provides it from the same placement table: posed_qpos returns the sampled qpos (the host sampler at seed 0, lane 0), and task_meta_words returns the tape, mask, init words and shaping words at the meta indices the hooks read. The SO-101 tower viewers call these, so a viewer cannot pose a scene the trainer would not.