`stage1_pickplace_as@v1` — Stage-1 pick-and-place — action-space (AS) control

A tabletop pick-and-place where the ego arm moves a cube to a goal while a second robot shares the cell, in a homogeneous panda/panda and a heterogeneous panda/fetch pairing. Per ADR-026 the task is ego-solvable, so its action-space cell is construct-invalid for cooperation claims. It is retained precisely as a control: a method claiming cooperation gains here is measuring something else.

Tier 1 — controls · Status: CONTROL · Suite: CHAMBER-Bench v1.0 (adr/ADR-027-chamber-bench-v1-protocol.md)

Spaces

Action space. Dict over ego + partner; homogeneous panda/panda pair vs heterogeneous panda/fetch pair (condition_id selects within the preregistered AS pair).
Observation. state_dict observations (qpos/qvel + task state); Stage1ASStateSynthesizer presents a synthesised flat state channel.
Simulator backend. maniskill3 · Agents. 2

Stress channel

None — no force coupling channel.

Axis validity

Per-cell status under the validity matrix (adr/ADR-027-chamber-bench-v1-protocol.md):

AS	OM	CR	CM	PF	SA
`invalid`	`untested`	`untested`	`untested`	`untested`	`untested`

Industrial analogue

A single-robot bin-to-conveyor pick-and-place station with a second robot present in the same cell — the everyday case where co-presence does not imply cooperation, which is exactly what the control checks.

Evidence

How to run one episode

import chamber.tasks

env = chamber.tasks.make("stage1_pickplace_as")
obs, info = env.reset(seed=0)
# Requires a SAPIEN-capable host (Tier-2).

Notes

Construct-invalid for cooperation as operationalized (ego-solvable; the AS cell is invalid per ADR-026's reinterpretation of the immutable 2026-06-11 archive) — retained precisely as a Tier-1 control (ADR-027 §Tier ladder): a method claiming cooperation gains here is measuring something else. CONTROL is now measured, not just recorded: the committed admission report (tag prereg-admission-stage1-pickplace-as-2026-07-05) rejects the task — A1 scripted-competent ego 60/60, A2 FAIL (partner-ablated 60/60: not two-robot-infeasible), A3 FAIL (partner-blind 60/60; gap CI [0, 0] < 0.20 — the demotion rule). Default condition is the heterogeneous panda+fetch cell; pass condition_id='stage1_pickplace_panda_only_mappo_shared_param' for the homogeneous baseline.

stage1_pickplace_as@v1 — Stage-1 pick-and-place — action-space (AS) control