The dataset viewer is not available for this subset.
Exception: SplitsNotFoundError
Message: The split names could not be parsed from the dataset config.
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
for split_generator in builder._split_generators(
~~~~~~~~~~~~~~~~~~~~~~~~~^
StreamingDownloadManager(base_path=builder.base_path, download_config=download_config)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/webdataset/webdataset.py", line 84, in _split_generators
raise ValueError(
...<2 lines>...
)
ValueError: The TAR archives of the dataset should be in WebDataset format, but the files in the archive don't share the same prefix or the same types.
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/split_names.py", line 68, in compute_split_names_from_streaming_response
for split in get_dataset_split_names(
~~~~~~~~~~~~~~~~~~~~~~~^
path=dataset,
^^^^^^^^^^^^^
config_name=config,
^^^^^^^^^^^^^^^^^^^
token=hf_token,
^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
info = get_dataset_config_info(
path,
...<6 lines>...
**config_kwargs,
)
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.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.
Daedalus × RoboCasa365 Target50 — full evaluation records
Complete per-episode records of two Daedalus agent configurations on the RoboCasa365 Target50 protocol (RPent's evaluation set: 18 atomic tasks × seeds 1–10 + 16 composite-seen × seeds 1–5 + 16 composite-unseen × seeds 1–5 = 340 episodes, split=target, layout 0). Later large-scale runs (seeds up to 50) will be added to this repository.
Results (340 episodes each)
| Configuration | Overall (task-weighted) | Episodes | Atomic | Composite-Seen | Composite-Unseen |
|---|---|---|---|---|---|
| E10a — native stage programs + SAM 3.1 object table | 73.0% | 265/340 (77.9%) | 165/180 (91.7%) | 54/80 (67.5%) | 46/80 (57.5%) |
| E7a3 — native stage programs (no perception tool) | 67.6% | 252/340 (74.1%) | 166/180 (92.2%) | 44/80 (55.0%) | 42/80 (52.5%) |
Overall = mean of the 50 per-task success rates (RPent leaderboard convention). For reference, RPent (GPT-6 Astra, high) reports 59.2% on the same protocol (Atomic 87.8%, Composite-Seen 43.8%, Composite-Unseen 42.5%).
Per task (success / episodes)
| Task | Category | E10a | E7a3 |
|---|---|---|---|
| CloseBlenderLid | atomic | 6/10 | 5/10 |
| CloseFridge | atomic | 10/10 | 10/10 |
| CloseToasterOvenDoor | atomic | 10/10 | 10/10 |
| CoffeeSetupMug | atomic | 8/10 | 9/10 |
| NavigateKitchen | atomic | 10/10 | 10/10 |
| OpenCabinet | atomic | 10/10 | 9/10 |
| OpenDrawer | atomic | 9/10 | 9/10 |
| OpenStandMixerHead | atomic | 8/10 | 9/10 |
| PickPlaceCounterToCabinet | atomic | 8/10 | 9/10 |
| PickPlaceCounterToStove | atomic | 10/10 | 10/10 |
| PickPlaceDrawerToCounter | atomic | 10/10 | 10/10 |
| PickPlaceSinkToCounter | atomic | 10/10 | 10/10 |
| PickPlaceToasterToCounter | atomic | 10/10 | 10/10 |
| SlideDishwasherRack | atomic | 10/10 | 10/10 |
| TurnOffStove | atomic | 9/10 | 10/10 |
| TurnOnElectricKettle | atomic | 10/10 | 10/10 |
| TurnOnMicrowave | atomic | 10/10 | 10/10 |
| TurnOnSinkFaucet | atomic | 7/10 | 6/10 |
| DeliverStraw | composite_seen | 5/5 | 4/5 |
| GetToastedBread | composite_seen | 5/5 | 4/5 |
| KettleBoiling | composite_seen | 4/5 | 5/5 |
| LoadDishwasher | composite_seen | 3/5 | 2/5 |
| PackIdenticalLunches | composite_seen | 0/5 | 0/5 |
| PreSoakPan | composite_seen | 5/5 | 4/5 |
| PrepareCoffee | composite_seen | 4/5 | 4/5 |
| RinseSinkBasin | composite_seen | 5/5 | 4/5 |
| ScrubCuttingBoard | composite_seen | 5/5 | 5/5 |
| SearingMeat | composite_seen | 3/5 | 2/5 |
| SetUpCuttingStation | composite_seen | 4/5 | 3/5 |
| StackBowlsCabinet | composite_seen | 3/5 | 2/5 |
| SteamInMicrowave | composite_seen | 1/5 | 0/5 |
| StirVegetables | composite_seen | 2/5 | 1/5 |
| StoreLeftoversInBowl | composite_seen | 0/5 | 0/5 |
| WashLettuce | composite_seen | 5/5 | 4/5 |
| ArrangeBreadBasket | composite_unseen | 4/5 | 5/5 |
| ArrangeTea | composite_unseen | 5/5 | 3/5 |
| BreadSelection | composite_unseen | 4/5 | 5/5 |
| CategorizeCondiments | composite_unseen | 2/5 | 3/5 |
| CuttingToolSelection | composite_unseen | 5/5 | 5/5 |
| GarnishPancake | composite_unseen | 1/5 | 0/5 |
| GatherTableware | composite_unseen | 3/5 | 2/5 |
| HeatKebabSandwich | composite_unseen | 2/5 | 1/5 |
| MakeIceLemonade | composite_unseen | 0/5 | 0/5 |
| PanTransfer | composite_unseen | 4/5 | 2/5 |
| PortionHotDogs | composite_unseen | 5/5 | 5/5 |
| RecycleBottlesByType | composite_unseen | 3/5 | 5/5 |
| SeparateFreezerRack | composite_unseen | 0/5 | 0/5 |
| WaffleReheat | composite_unseen | 2/5 | 1/5 |
| WashFruitColander | composite_unseen | 2/5 | 2/5 |
| WeighIngredients | composite_unseen | 4/5 | 3/5 |
Protocol and conditions
- Planner: Codex with gpt-6-astra, reasoning effort high; at most 100 completed model requests per episode; no wall-clock or simulator-step limit; one run per cell.
- Control: Daedalus Phase B native stage programs for the PandaOmron mobile manipulator (no VLA). Observations: RGB + proprioception + RPent depth-derived world maps (
back_project_batch,query_world_map) and RPent's task predicates (success_criteria,task_progress) — the same information RPent's planner receives. No simulator object poses. E10a additionally exposes anobject_tabletool (SAM 3.1 text segmentation lifted to 3-D on the world map). - Note this is RPent's Target50 protocol, not the official RoboCasa365 leaderboard protocol (pretrain kitchens, step horizons, no in-episode progress signal).
- API relay: E10a — 306 episodes on the primary relay, 34 (seed 1) on micuapi; 16 seed-1 episodes cut off by a relay quota error were re-run and the cut-off attempts are kept in
runs_superseded/. E7a3 — micuapi throughout (Codex memory model aliased gpt-5.6-terra→gpt-6-sol). - Code: E10a = Daedalus commit
9d43d6c(branch robocasa/e10); E7a3 = commit52e54c6(branch robocasa/e6b, E6 switches off). The exact code tree is inside each archive undercode/.
Files
| File | Content |
|---|---|
daedalus-robocasa365-target50-E10a-sam-astra-9d43d6c-20261030.tar.zst.part00–part01 |
E10a archive (~17.6 GB) |
daedalus-robocasa365-target50-E7a3-astra-52e54c6-20261029.tar.zst.part00–part01 |
E7a3 archive (~17.0 GB) |
daedalus-robocasa365-target50-E10a-full2500-official-astra-9d43d6c-20261104.tar.zst.part00–part12 |
E10a full 2500-episode archive, official attempts (~131 GB) |
SHA256SUMS |
checksums of every part |
Each archive contains runs/<category>/<Task>_s<seed>/ (summary.json, codex-raw.jsonl, events.sqlite, effective_config.json, blobs/, scene/), results/cells.csv (one row per episode), metadata/ (queue manifests, configs, registry entries), queues/, code/, README.txt and report.md.
Extract:
sha256sum -c --ignore-missing SHA256SUMS
cat daedalus-robocasa365-target50-E10a-sam-astra-*.tar.zst.part* | zstd -d | tar -xf -
Secrets: every auth.json (local proxy placeholder key) was excluded; the uncompressed archive streams were scanned for the real API keys and the HF token with zero matches.
Full run: Target50 × seeds 1–50 (2500 episodes, E10a)
The E10a configuration above (commit 9d43d6c, gpt-6-astra with reasoning effort high, at most 100 completed model requests per episode, no wall-clock or simulator-step limit) on all 50 tasks × seeds 1–50: the 340 E10a episodes above plus 2160 new ones. Every task has 50 episodes, so task-weighted and per-episode rates are equal.
| Scoring | Overall | Atomic | Composite-Seen | Composite-Unseen |
|---|---|---|---|---|
| Official | 71.8% (1794/2500; 95% CI 70.4–73.0) | 92.9% (836/900) | 60.5% (484/800) | 59.2% (474/800) |
| First scored attempt only | 71.3% (1783/2500; 95% CI 70.0–72.6) | 92.9% (836/900) | 59.1% (473/800) | 59.2% (474/800) |
Quote both rows. The rule fixed before the run was "the first scored attempt of each cell counts". Composite-seen success dipped at seeds 6–12, so those 98 cells (14 tasks × seeds 6–12) were rerun. After seeing the rerun (55 → 66 successes), the experimenter chose the rerun as the official score, judging the dip to be variation in the upstream model rather than an infrastructure fault; the logs can neither confirm nor rule this out. Because these cells were selected for being low, the official row is optimistic; the first-attempt row follows the rule fixed in advance.
- Intervals come from resampling seeds with the tasks fixed; they do not cover other tasks, model versions or API conditions.
- One cell (SlideDishwasherRack_s26) never produced a scored result (all three attempts stopped with
base_calibration_degenerate) and counts as a failure. - How the 2500 official episodes ended: native episode end 1794, actor made no progress 412, model request limit 293, harness stop 1.
- Mean completed model requests per episode: atomic 30.1, composite-seen 56.5, composite-unseen 56.0.
- Under RPent's wall-clock limits (atomic 1800 s, composite 3600 s) the official result would be 67.1% (first attempt 66.5%).
Per task, 2500-episode run (successes / 50 episodes)
| Task | Category | Official | First attempt |
|---|---|---|---|
| CloseBlenderLid | atomic | 32/50 | 32/50 |
| CloseFridge | atomic | 50/50 | 50/50 |
| CloseToasterOvenDoor | atomic | 50/50 | 50/50 |
| CoffeeSetupMug | atomic | 39/50 | 39/50 |
| NavigateKitchen | atomic | 50/50 | 50/50 |
| OpenCabinet | atomic | 50/50 | 50/50 |
| OpenDrawer | atomic | 49/50 | 49/50 |
| OpenStandMixerHead | atomic | 44/50 | 44/50 |
| PickPlaceCounterToCabinet | atomic | 42/50 | 42/50 |
| PickPlaceCounterToStove | atomic | 50/50 | 50/50 |
| PickPlaceDrawerToCounter | atomic | 46/50 | 46/50 |
| PickPlaceSinkToCounter | atomic | 49/50 | 49/50 |
| PickPlaceToasterToCounter | atomic | 50/50 | 50/50 |
| SlideDishwasherRack | atomic | 48/50 | 48/50 |
| TurnOffStove | atomic | 49/50 | 49/50 |
| TurnOnElectricKettle | atomic | 50/50 | 50/50 |
| TurnOnMicrowave | atomic | 50/50 | 50/50 |
| TurnOnSinkFaucet | atomic | 38/50 | 38/50 |
| DeliverStraw | composite_seen | 35/50 | 33/50 |
| GetToastedBread | composite_seen | 40/50 | 39/50 |
| KettleBoiling | composite_seen | 48/50 | 48/50 |
| LoadDishwasher | composite_seen | 26/50 | 24/50 |
| PackIdenticalLunches | composite_seen | 4/50 | 4/50 |
| PreSoakPan | composite_seen | 36/50 | 38/50 |
| PrepareCoffee | composite_seen | 33/50 | 33/50 |
| RinseSinkBasin | composite_seen | 41/50 | 41/50 |
| ScrubCuttingBoard | composite_seen | 50/50 | 50/50 |
| SearingMeat | composite_seen | 23/50 | 20/50 |
| SetUpCuttingStation | composite_seen | 43/50 | 41/50 |
| StackBowlsCabinet | composite_seen | 25/50 | 24/50 |
| SteamInMicrowave | composite_seen | 17/50 | 18/50 |
| StirVegetables | composite_seen | 28/50 | 27/50 |
| StoreLeftoversInBowl | composite_seen | 3/50 | 3/50 |
| WashLettuce | composite_seen | 32/50 | 30/50 |
| ArrangeBreadBasket | composite_unseen | 31/50 | 31/50 |
| ArrangeTea | composite_unseen | 44/50 | 44/50 |
| BreadSelection | composite_unseen | 47/50 | 47/50 |
| CategorizeCondiments | composite_unseen | 36/50 | 36/50 |
| CuttingToolSelection | composite_unseen | 49/50 | 49/50 |
| GarnishPancake | composite_unseen | 3/50 | 3/50 |
| GatherTableware | composite_unseen | 30/50 | 30/50 |
| HeatKebabSandwich | composite_unseen | 13/50 | 13/50 |
| MakeIceLemonade | composite_unseen | 5/50 | 5/50 |
| PanTransfer | composite_unseen | 40/50 | 40/50 |
| PortionHotDogs | composite_unseen | 47/50 | 47/50 |
| RecycleBottlesByType | composite_unseen | 46/50 | 46/50 |
| SeparateFreezerRack | composite_unseen | 0/50 | 0/50 |
| WaffleReheat | composite_unseen | 18/50 | 18/50 |
| WashFruitColander | composite_unseen | 25/50 | 25/50 |
| WeighIngredients | composite_unseen | 40/50 | 40/50 |
Files of the 2500-episode run
daedalus-robocasa365-target50-E10a-full2500-official-astra-9d43d6c-20261104.tar.zst.part00–part12 (~131 GB in all). For every cell the archive holds the attempt that gives the official score, in runs/<category>/<Task>_s<seed>/ (summary.json, codex-raw.jsonl, events.sqlite, effective_config.json, blobs/, scene/, private/). It also holds results/cells.csv (one row per cell, with both the official and the first-attempt outcome), results/summary.json, report.md (Chinese), metadata/ (queue manifests and launch scripts of every batch, condition configs, the list of the 98 replaced cells, registry entries), code/ (the exact code tree) and README.txt. The replaced first attempts of the 98 cells are not included. Every auth.json (local proxy placeholder key) and the events.sqlite-wal/-shm side files were left out, and the uncompressed stream was scanned for the real API keys and the HF token with zero matches.
sha256sum -c daedalus-robocasa365-target50-E10a-full2500-official-astra-9d43d6c-20261104.tar.zst.sha256
cat daedalus-robocasa365-target50-E10a-full2500-official-astra-9d43d6c-20261104.tar.zst.part* | zstd -d | tar -xf -
- Downloads last month
- 122