alexmzhang/castle-reasoning

Castle construction demonstrations, 2026-10-10 to 2026-10-11

1,648 accepted demonstrations of 1,072 different construction targets. In each, an OGBench UR5e with a Robotiq 2F-85 gripper builds a procedurally generated target out of cuboid parts in MJ-Warp. Two campaigns collected them on one RTX 5090.

Code: https://github.com/smbml/castle-reasoning-benchmark, branch robo-alex/campaign-14-results, commit c168619. The write-up is docs/construction-campaign.md; the commands and the evidence are in docs/benchmarks/construction-campaign-20261010/. The collector code at c168619 is the code both campaigns ran (the first campaign's driver started at 15b8275 and was restarted at dea0043; only the campaign driver differs between the two).

Layout

This repository mirrors the artifacts/ directory of the checkout that collected the data. Download it into artifacts/ of a checkout of the branch:

hf download alexmzhang/castle-reasoning --repo-type dataset --local-dir artifacts

Path	Contents
campaign-1k-20261010/	First campaign: 1,146 demonstrations of 570 targets (up to two per target), seed 1000000000
campaign-tasks-20261011/	Second campaign: 502 demonstrations of 502 other targets (one each), seed 100000000
scale-probe-20261009/pool1000-slots/plans/	The first campaign's 579 plans (S0001 ...)
scale-probe-20261009/screen579-rep2/summary.json, scale-probe-20261009/screen-slots/summary.json	Stability screens of those plans
tasks-20261011/pool/plans/, tasks-20261011/screen/summary.json	The second campaign's 533 plans (U0001 ...) and their screen
SHA256SUMS	Checksums of every other file here; check with sha256sum -c SHA256SUMS

In each campaign directory:

• campaign.json: the driver settings, the collector flags, and every plan and screen file with its SHA-256. Its paths are absolute paths on the collecting machine (/home/alexzhang/castle-reasoning-benchmark/...).
• index.json: one entry per accepted demonstration: target_id, fingerprint, parts, fixed_parts, physics_timestep, shard_id, run_group, run, seed, attempt_directory (relative to the campaign directory) and action_targets (number of actions).
• shards/<shard>/ and topups/<run group>/: runs.json, and per collector run runNN/summary.json and runNN.log. The summaries list every attempt, including failed ones; only accepted attempts are uploaded.

In each accepted attempt directory:

File	Contents
replay.npz	The demonstration: its recorded actions replayed from a fresh reset of the parked start state. action (T, 7) float32 at 50 Hz, six normalized arm joint velocities and an absolute gripper command; grip_effort (T); qpos (T+1, nq), qvel (T+1, nv) and time (T+1).
model.mjb	The compiled MuJoCo model it ran on (mujoco 3.12.0). Models are padded to the run's part slots; slots at or past plan.json's active_part_count hold unused parts parked off the table.
plan.json	The target, the removal and assembly orders, goals and staging.
execution_config.json, execution_order.json	Collector settings and the executed order.
provenance.json	Package versions, third-party commits and SHA-256 of the code.
settle_checks.json, disassembly_checks.json, assembly_checks.json, replay_checks.json	The acceptance checks it passed.
replay_settle.npz	Only in some attempts: the extra hold after the replay used for verification. It is not part of the demonstration.

Left out to save space: the physics traces of the forward disassembly and assembly (disassembly.npz, assembly.npz, settle.npz), the segment logs, failed attempts and the collector's working model copies.

Load a demonstration

import mujoco
import numpy as np

attempt = "campaign-tasks-20261011/shards/2ms-b0-000/run00/attempt00150"
model = mujoco.MjModel.from_binary_path(f"{attempt}/model.mjb")
data = mujoco.MjData(model)
demo = np.load(f"{attempt}/replay.npz")
data.qpos[:], data.qvel[:] = demo["qpos"][-1], demo["qvel"][-1]
mujoco.mj_forward(model, data)  # the finished building

The actions are commands to the collector's controller, not MuJoCo ctrl values. Simulating them again uses the collector's replay (Collection.replay in robot_assembly/continuous_collect.py) on the MJ-Warp runtime.

Collect the data again

Set up the environment as in the code repository's CLAUDE.md and docs/construction_setup.md (uv sync, build the active-world runtime into .runtime-active, then python -m scripts.check_construction_install --require-active). From the checkout root, with this repository in artifacts/:

export PYTHONPATH="$PWD/.runtime-active:$PWD" MUJOCO_GL=egl

# First campaign
uv run python -m scripts.run_construction_campaign \
  --plans artifacts/scale-probe-20261009/pool1000-slots/plans/*.json \
  --output OUT1 \
  --screen artifacts/scale-probe-20261009/screen579-rep2/summary.json \
           artifacts/scale-probe-20261009/screen-slots/summary.json \
  --repeats 2 --seed 1000000000

# Second campaign
uv run python -m scripts.run_construction_campaign \
  --plans artifacts/tasks-20261011/pool/plans/*.json \
  --output OUT2 --screen artifacts/tasks-20261011/screen/summary.json \
  --repeats 1 --max-rounds 1 --seed 100000000 \
  --max-plans-per-shard 244 --min-plans-per-shard 100

The uploaded screens let you skip screening; to screen again, run python -m scripts.screen_construction_stability --plans PLANS --output SCREEN (seed 20261010 for the first pool, 20261011 for the second).

For comparison: the first campaign accepted 1,146 of 1,186 attempts (570 of 573 targets) in 14.5 hours, the second 502 of 527 in 7.1 hours, on one RTX 5090 with 32 CPU threads. Whether a rerun reproduces individual trajectories exactly has not been tested: planning has wall-clock budgets and 256 worlds share the GPU and CPUs, so timing can change outcomes. Compare acceptance per target.