Dataset Viewer
Duplicate
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
experiment: string
label: string
strategy: string
agents: int64
group: string
seed: int64
timeoutSec: int64
wallMs: int64
cpuMs: int64
outcome: string
interactive: bool
durationSec: double
model: string
antsPerAgent: int64
ants: int64
sites: int64
minPath: double
pathLength: double
pathRatio: double
initialDistance: double
finalDistance: double
bestDistance: double
progress: double
state: string
states: string
firstChoice: string
bEntries: int64
stateTime: string
slit1Sec: double
grips: int64
squeezed: int64
meanAttached: double
meanPulling: double
meanInformed: double
touching: double
stateShare: struct<walk: double, rest: double, lead: double, pull: double, hold: double>
  child 0, walk: double
  child 1, rest: double
  child 2, lead: double
  child 3, pull: double
  child 4, hold: double
hl: struct<name: string, title: string, hash: string, version: int64, id: string, errors: int64>
  child 0, name: string
  child 1, title: string
  child 2, hash: string
  child 3, version: int64
  child 4, id: string
  child 5, errors: int64
hlChampion: struct<id: string, name: string, version: int64, file: string>
  child 0, id: string
  child 1, name: string
  child 2, version: int64
  child 3, file: string
to
{'experiment': Value('string'), 'label': Value('string'), 'strategy': Value('string'), 'agents': Value('int64'), 'group': Value('string'), 'seed': Value('int64'), 'timeoutSec': Value('int64'), 'wallMs': Value('int64'), 'cpuMs': Value('int64'), 'outcome': Value('string'), 'interactive': Value('bool'), 'durationSec': Value('float64'), 'model': Value('string'), 'antsPerAgent': Value('int64'), 'ants': Value('int64'), 'sites': Value('int64'), 'minPath': Value('float64'), 'pathLength': Value('float64'), 'pathRatio': Value('float64'), 'initialDistance': Value('float64'), 'finalDistance': Value('float64'), 'bestDistance': Value('float64'), 'progress': Value('float64'), 'state': Value('string'), 'states': Value('string'), 'firstChoice': Value('string'), 'bEntries': Value('int64'), 'stateTime': Json(decode=True), 'slit1Sec': Value('float64'), 'grips': Value('int64'), 'squeezed': Value('int64'), 'meanAttached': Value('float64'), 'meanPulling': Value('float64'), 'meanInformed': Value('float64'), 'touching': Value('float64'), 'stateShare': {'walk': Value('float64'), 'rest': Value('float64'), 'lead': Value('float64'), 'pull': Value('float64'), 'hold': Value('float64')}}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
                  return get_rows(
                      dataset=dataset,
                  ...<4 lines>...
                      column_names=column_names,
                  )
                File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
                  return func(*args, **kwargs)
                File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
                  rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
                File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
                  yield from ds.decode(False) if ds.features else ds
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2951, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2461, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2486, in _iter_arrow
                  for key, pa_table in self.ex_iterable._iter_arrow():
                                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 547, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 430, in _iter_arrow
                  for key, pa_table in self.generate_tables_fn(**gen_kwags):
                                       ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              experiment: string
              label: string
              strategy: string
              agents: int64
              group: string
              seed: int64
              timeoutSec: int64
              wallMs: int64
              cpuMs: int64
              outcome: string
              interactive: bool
              durationSec: double
              model: string
              antsPerAgent: int64
              ants: int64
              sites: int64
              minPath: double
              pathLength: double
              pathRatio: double
              initialDistance: double
              finalDistance: double
              bestDistance: double
              progress: double
              state: string
              states: string
              firstChoice: string
              bEntries: int64
              stateTime: string
              slit1Sec: double
              grips: int64
              squeezed: int64
              meanAttached: double
              meanPulling: double
              meanInformed: double
              touching: double
              stateShare: struct<walk: double, rest: double, lead: double, pull: double, hold: double>
                child 0, walk: double
                child 1, rest: double
                child 2, lead: double
                child 3, pull: double
                child 4, hold: double
              hl: struct<name: string, title: string, hash: string, version: int64, id: string, errors: int64>
                child 0, name: string
                child 1, title: string
                child 2, hash: string
                child 3, version: int64
                child 4, id: string
                child 5, errors: int64
              hlChampion: struct<id: string, name: string, version: int64, file: string>
                child 0, id: string
                child 1, name: string
                child 2, version: int64
                child 3, file: string
              to
              {'experiment': Value('string'), 'label': Value('string'), 'strategy': Value('string'), 'agents': Value('int64'), 'group': Value('string'), 'seed': Value('int64'), 'timeoutSec': Value('int64'), 'wallMs': Value('int64'), 'cpuMs': Value('int64'), 'outcome': Value('string'), 'interactive': Value('bool'), 'durationSec': Value('float64'), 'model': Value('string'), 'antsPerAgent': Value('int64'), 'ants': Value('int64'), 'sites': Value('int64'), 'minPath': Value('float64'), 'pathLength': Value('float64'), 'pathRatio': Value('float64'), 'initialDistance': Value('float64'), 'finalDistance': Value('float64'), 'bestDistance': Value('float64'), 'progress': Value('float64'), 'state': Value('string'), 'states': Value('string'), 'firstChoice': Value('string'), 'bEntries': Value('int64'), 'stateTime': Json(decode=True), 'slit1Sec': Value('float64'), 'grips': Value('int64'), 'squeezed': Value('int64'), 'meanAttached': Value('float64'), 'meanPulling': Value('float64'), 'meanInformed': Value('float64'), 'touching': Value('float64'), 'stateShare': {'walk': Value('float64'), 'rest': Value('float64'), 'lead': Value('float64'), 'pull': Value('float64'), 'hold': Value('float64')}}
              because column names don't match

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

HEMPO: Hierarchical Embodied Multi-Agent with Partial Observation

A cooperative multi-agent environment built around the "piano movers" geometric puzzle from the PNAS paper Comparing cooperative geometric puzzle solving in ants versus humans (Dreyer et al., DOI: 10.1073/pnas.2414274121). A T-shaped load must be carried through three chambers connected by openings narrower than the load's long crossbar β€” it must be turned, threaded and rolled over, and no single agent is strong enough to move it alone.

The repository holds three layers of the same science:

  1. mc-collective-transport/ β€” the environment and experiments: three cooperative-transport tasks in Minecraft and in a fast headless simulator, driven by a self-improving heuristic framework whose top-level "researcher" is an LLM.
  2. photoenv/ β€” the same environments rebuilt photo-real in Blender, with every dimension traced to the planners' own output (stills, clips and the geometry pipeline).
  3. mc-bots/ + mcserver/ β€” the original HEMPO Minecraft build: one leader scheduler and four worker bots solving the puzzle on a real Minecraft server.

The experiments (mc-collective-transport/)

task what the agents must do
transport box pushing with local perception (Chen et al., IEEE T-RO 2015): six agents push a box that needs three to move, toward a goal the box hides
piano the PNAS piano-movers puzzle: ~80 ants' worth of carriers move the T-load through two 5.32-block slits (paper geometry and the Table S5 ant model)
piano3d the load flown by flying agents through a horizontal window, rolled over in a short middle room, then through an upright window (6-DOF, reference optimum included)

Every task is scored against the global optimum from the repository's own configuration-space planner. On top of this runs the hl framework (agents/src/hl/): a three-layer loop β€” an LLM researcher (scheduled Claude task) proposes decision rules, an engine tunes them with CMA-ES and promotes champions, and the winning rules run on the agents in Minecraft. See mc-collective-transport/docs/hl-framework.md.

colony-llm: LLM agents benchmark

mc-collective-transport/agents/src/colony/ turns the ant task into a benchmark where each ant is an agent whose decisions come from an LLM (any OpenAI-compatible endpoint, or a local OpenJev "System One" decision server). Baselines: the planner optimum, the paper's stochastic ant model, and a memoryless control. Headline result at 80 ants / 2500 s (report):

policy success note
planner (global optimum) 1/1 in 37 s path ratio 1.05
paper ant model 3/3 (629–1557 s) stochastic self-organisation
deepseek-v4.1-flash / glm-5-3-flash / mimo-v2.6-pro 0/3 deadlock at slit 1
OpenJev (Laya, System One) 0/3 all-pull equilibrium

The LLM ants reproduce the colony's labour division yet none solve the task: the ant model wins by continuously re-sampling configurations, not by better single decisions. Environment spec and reproduction commands: docs/colony-llm.md; live dashboard: node agents/src/colony/live.js (http://127.0.0.1:25610).

human-llm: the human condition and its coordination mechanisms

--human --mech ... turns the same puzzle into the paper's human condition: 16 people, one handle each, exclusively pulling, every person an LLM agent fed its own observation history. The four mechanisms differ only in the coordination channel: solo (no communication), comm (shared chat), chief-local (an informed-leader chief who walks the arena floor and has no global view), and chief-god (the chief sees the whole arena). The chief is one of the people β€” same model, never pulls.

The benchmark's real-time protocol (--sync) makes the world wait while any person or the chief is waiting for its LLM reply, so every action follows a fresh decision. Over 10 seeds Γ— 2500 s (report):

mechanism solved fastest solve mean pulling
chief-local 1/10 640.8 s 11.3/16
chief-god 6/10 316.9 s 15.5/16

For comparison, the earlier unpaced protocol scored 0/10 and 3/10 (fastest 2057 s). Spec, prompts and commands: docs/human-llm.md; per-seed results in agents/data/colony/results/, recorded clips in photoenv/videos/.

Photo-real rebuild (photoenv/)

The environment as a physical facility, rebuilt from scratch in Blender β€” nothing is eyeballed: scripts/plan_path.js / plan_path3.js run the repository's own planners and every wall, slit, load contour and manoeuvre in the renders comes from that dump. Three scenes: the robot hall (humanoid carriers, the box-pushing arena beside the puzzle), the ant colony at true ant scale (134 procedural Paratrechina longicornis around the red PLA rim load, per the paper's Materials and Methods), and the 3D flight bay (hexacopter carriers threading the two windows along the saved 6-DOF optimum). Deliverables, fidelity table and commands: photoenv/README.md; the visual pipeline is documented in photoenv/HANDOVER.md.

Live views and recordings

The simulation can be watched live instead of rendered offline. photoenv/unity-live-proj/ is a Unity 6 project that streams the running trial's pose (/tmp/live_web/state.json) into an interactive top-down scene, with four first-person panels and the chat panel on the side; photoenv/scripts/live_web.py serves the same scene as a Three.js page for a browser. The 27 recorded clips β€” m3/m4 Γ— 10 seeds, plus the successful --sync runs β€” are in photoenv/videos/.

The original HEMPO build (mc-bots/, mcserver/)

  • mc-bots/ β€” Node.js (mineflayer) bot system
    • leader.js β€” central scheduler: DFS over the 8-node configuration space, broken-edge long-term memory, multi-agent coordination
    • worker_pull.js β€” collaborative push/pull control for the boat-carried load (boat + leash-knot handles)
    • build_scene.js / clear_environment.js β€” scene builder and cleanup
    • README.md β€” puzzle design and algorithm details
  • mcserver/ β€” Minecraft Java server configuration (world saves, jars and other runtime files are not included)

Running: start a Minecraft server (Java 21, 1.21.x) in mcserver/ on localhost:25565, then cd mc-bots && npm install and npm run puzzle for the full pipeline.

Repository layout

mc-collective-transport/   environment + experiments (Node.js)
  agents/src/{core,hl,sim,mc,colony,llm}   physics, planners, hl framework, colony-llm benchmark
  docs/                    hl-framework.md, colony-llm.md, human-llm.md
  scripts/                 colony-env.sh, colony-benchmark.sh, human-benchmark.sh
  code/                    experiment snapshots and reports
photoenv/                  photo-real Blender rebuild (Python/Node scripts + renders)
  unity-live-proj/         Unity 6 live view (streams the running trial's pose)
  videos/                  recorded benchmark clips (m3/m4 x 10 seeds + --sync runs)
mc-bots/                   original mineflayer bot system
mcserver/                  Minecraft server configuration

Citation

Dreyer, T., et al. "Comparing cooperative geometric puzzle solving in ants versus humans." PNAS. DOI: 10.1073/pnas.2414274121

Downloads last month
126