`cocarry@v1` — Co-carry — rigid dual-arm bar transport

Two manipulators rigidly carry one shared bar to a goal while keeping it level and keeping internal stress within a frozen force ceiling. The task is infeasible for one robot, so success measures cooperation rather than a single ego's competence. Its rung archives (frozen incumbent, policy-shift and embodiment-shift measurements) are the committed evidence for Tier-2 admission.

Tier 2 — admitted cooperation tasks · Status: ADMITTED · Suite: CHAMBER-Bench v1.0 (adr/ADR-027-chamber-bench-v1-protocol.md)

Spaces

Action space. Dict over ego (panda_wristcam) + partner (panda_partner, or xarm6_robotiq under the embodiment-shift condition); per-agent arm control with a rigid dual-hold bar attach.
Observation. State observations including bar pose, tilt, and wrist constraint-solver force telemetry.
Simulator backend. maniskill3 · Agents. 2

Stress channel

Wrist constraint-solver force — the 6-axis spatial force through each holding arm's hand incoming joint — as it gates success on main (canonical instrument; f_max = 130.5697 N per spikes/results/cocarry/rung2/cocarry_rung2_freeze_manifest.json). The rung-4c embodiment-invariant bar coupling-stress instrument (365.6 N ceiling, post-hoc only) is secondary telemetry per ADR-027 §Versioning's canonical-instrument rule.

Axis validity

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

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

Industrial analogue

Dual-robot handling of long or heavy payloads — two industrial arms sharing one rigid part, as in dual-arm transport of sheet-metal or chassis components in automotive body shops, where tilt and internal stress limits bind exactly because two holds share the load.

Evidence

How to run one episode

import chamber.tasks

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

Notes

ADMITTED by the committed admission report (ADR-027 §Admission protocol; tag prereg-admission-cocarry-2026-07-05): A1 scripted matched reference 60/60 with wrist stress max 107.9 N < f_max 130.5697; A2 single-arm positive control 0/60; A3 partner-blind ego (B-BLIND, cocarry_blind_impedance) 0/60 — paired gap CI [1.0, 1.0] >= delta_min 0.20. Campaign finding (2026-07-06, spikes/results/benchmark/cocarry-v1/CAMPAIGN_REPORT.md): the LEARNED B-BLIND-vs-B-AHT contrast is an informative null (both 1.000 — masked partner state re-enters through the ego's own proprioception under rigid coupling; ADR-027 §Revision 2026-07-06); discrimination lives in the stress channel (B-JOINT pairs ~41 N vs ~102-108 N cross-play) and the per-partner breakdown (imp_lag_bounded floor 0.88). Binding evidence: stress p50 88.6 / p99 106.4 N on matched successes. The null AS cell records the embodiment-heterogeneity robust null established by the rung-4 archives (rung-4d pose-artifact retraction + rung-4e task-fair pose/controller search); the rung-3 policy-heterogeneity pooled null (with its stiff-impedance caveat) is recorded in the rung3 archive. Changing the canonical success predicate or stress instrument creates cocarry@2 (ADR-027 §Versioning).

cocarry@v1 — Co-carry — rigid dual-arm bar transport