Skip to content

coinsert@v1 — Co-insert — hold-and-insert (closed; open challenge)

A two-robot hold-and-insert: the ego drives a peg into a blind socket carried by a free receptacle that only the partner stabilises, so a single inserter has nothing to push against. The task closed at a pre-committed HARD_STOP: no reference pair seats the peg at the committed tolerances because insertion contact cocks the peg past the seatable tilt region (a geometric tilt-wedge). The env stays wired and runnable as a frontier challenge task.

Tier 3 — documented candidates and closures · Status: CLOSED · Suite: CHAMBER-Bench v1.0 (adr/ADR-027-chamber-bench-v1-protocol.md)

Spaces

  • Action space. Dict over ego inserter (panda_wristcam) + holder (panda_partner gripping a free receptacle); per-agent arm control.
  • Observation. State observations including peg seating depth, relative tilt, and interaction-wrench telemetry.
  • Simulator backend. maniskill3 · Agents. 2

Stress channel

Insertion and coupling force limits (f_insert_max / f_couple_max), derived from measured matched-pair distributions

Axis validity

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

AS OM CR CM PF SA
untested untested untested untested untested untested

Industrial analogue

Two-robot assembly of close-tolerance connectors where one robot holds an unfixtured housing and the other inserts — the hardest-case cousin of standard robotic peg-in-hole insertion, here at a clearance regime that remains unsolved for the matched pair.

Open challenge

No reference pair has met the A1 solvability bar at the committed tolerances; the cause is understood (geometric tilt-wedge; the simulator-fidelity suspicion was tested against a MuJoCo oracle and disproven, #257-#258). Solving the matched-pair problem is the open challenge. No leaderboard cells exist for this task.

Evidence

How to run one episode

import chamber.tasks

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

Notes

Unsolved: no reference pair has met A1 at the committed tolerances — see the closure archive. The cause is understood (geometric tilt-wedge: the seatable region, tilt < 0.7°, and the achievable-control region under contact, ~0.9-2.8°, do not overlap); the simulator-fidelity suspicion was tested against a MuJoCo oracle and disproven (#257-#258). Solving the matched-pair problem is the open challenge; the env factory stays wired so the task is runnable as a frontier/challenge task. No leaderboard cells exist for this task.