ddffwyb/EchoPolicy-0-MolmoSpaces-eval

EchoPolicy-0: MolmoSpaces Evaluation

Evaluation results, complete episode videos, trajectories and execution logs for EchoPolicy-0, developed by MagicLab.

• Official submission: allenai/molmospaces#199
• Evaluation version: echopolicy-ms-v15-20260918-full3915
• Policy version: echo-tools-20260917-v15
• Coverage: 3915 episodes across Close, Pick, Open and Pick & Place, including every success and failure.

Browse the videos

Open the episode viewer.

Each row represents one episode. The exo_video and wrist_video columns show the two camera recordings side by side, alongside the task, seed, executed steps, episode horizon and official success labels.

The dataset contains 7830 original MP4 recordings, with both camera views for every episode. Videos preserve the full recorded episode, including early termination. The episode table is stored at videos/episodes.parquet; its video columns reference the original MP4 files at a fixed repository revision.

Results

Task	Episodes	Successes	At-end	Oracle	Official 95% interval
Close-v1	915	805	87.98%	87.98%	85.71–89.93%
Pick-v1.1	1000	810	81.00%	81.00%	78.45–83.31%
Open-v1	1000	671	67.10%	67.10%	64.13–69.94%
Pick & Place-v1	1000	424	42.40%	42.40%	39.37–45.49%

The mean success rate, with equal weight assigned to each of the four tasks, is 69.6195%. Success labels and confidence intervals come from the official evaluator and CSV exporter.

Method and evaluation protocol

EchoPolicy-0 combines a VLM Agent, Pi_05 and RGB-D geometric tools. The Agent observes the scene, selects and verifies subgoals, and routes execution to independent Close, Pick, Open and Place tools or to Pi_05. Visual verification determines whether to continue, retry or fall back. Pick-to-Place transitions require verified object holding.

• Robot: single-arm franka_droid with a Robotiq gripper. Actions contain seven absolute joint positions and one gripper command.
• Policy inputs: geometric tools use RGB, aligned exo metric depth, camera calibration and robot state. Pi_05 uses RGB and robot state. Simulator object poses, ground-truth masks and evaluator success labels are excluded from policy inputs.
• VLA scheduling: 16 candidate action sequences, each predicting 16 steps. Execute the selected sequence's first 8 steps, then predict again from a new observation. Geometric trajectories execute in chunks of at most 8 steps.
• Verification: routine VLM checks every 16 executed environment steps, plus checks at tool completion or failure. Tools allow up to three attempts per executor/subgoal; uncertain verification does not count as success.
• VLM: gemini-3.8-flash, thinking level low, without an explicit output-token cap or global/batch VLM concurrency cap.
• Official resources: 20260408; manifest hashes are recorded in metadata/run-manifest.json. This release uses the resource label Pick-v1.1; the leaderboard labels the task Pick-v1.
• Environment budget: policy_dt_ms=66 and end_on_success=True. The official evaluator derives horizons from manifest durations: 30 seconds for Close, Pick and Open (455 actions), and 45 seconds for Pick & Place (682 actions).
• Deployment: eight L40S GPUs, each hosting one Pi_05/Echo service and ten simulation workers. Workers use their local policy service; Pi_05 memory preallocation is 0.4 and model execution is serialized per service.
• Completeness: all 1767 batches completed on their first attempt. All outcomes were retained. The action audit covers 991,807 finite 7+1 actions.

The official CSVs were generated with the frozen exporter using --success-condition both --dt 0.066. The CSVs and export logs are available in results/.

Repository contents

videos/
  episodes.parquet               Episode table with both Video columns
  {task}/{manifest}/{episode}_exo.mp4
  {task}/{manifest}/{episode}_wrist.mp4
results/                         Official CSVs, CSV ZIP and export logs
archives/                        25 trajectory and execution-log archives
metadata/
  episodes.json                  Episode-to-video, trajectory and log mappings
  episodes.csv                   Flat episode index
  video-files.json               MP4 paths, sizes and SHA-256 hashes
  archives.json                  Archive paths, sizes and SHA-256 hashes
  *.members.json                 Original file paths and member hashes
  full-media-validation.json     Video decoding and frame-count verification
  vlm-request-index.jsonl        VLM request-to-episode mappings
  reproducibility/               Dependency, runtime and source-version records
scripts/
  verify_package.py              Verify payload and archive integrity
  extract_episode.py             Copy both videos and extract the batch HDF5
SHA256SUMS                       Payload file checksums

place in file paths denotes Pick & Place. metadata/episodes.json connects each episode ID to its seed, house, source manifest, environment steps, official outcome, both videos and the original HDF5 trajectory group. The source_path field records the original video filename and directory.

The 1767 HDF5 files preserve the official house/batch organization. Trajectories include simulator state and evaluator labels as evaluation outputs. Evaluation-log archives contain episode metadata, evaluator logs, batch manifests and VLA selection records. Orchestrator archives contain Agent events and geometric-tool inputs; VLM trace archives contain prompts, image payloads, generation settings and responses.

All 3915 formal episodes have saved VLM traces. The collection contains 74,860 request records, including service-level requests. Use the episode mappings in metadata/vlm-request-index.jsonl and coverage information in metadata/vlm-request-coverage.json when analyzing them.

Download and inspect

Install the Hugging Face CLI:

pip install -U huggingface_hub

Download the full dataset:

hf download ddffwyb/EchoPolicy-0-MolmoSpaces-eval \
  --repo-type dataset --local-dir ./echopolicy-eval

Download only the videos, episode table and video indexes:

hf download ddffwyb/EchoPolicy-0-MolmoSpaces-eval \
  --repo-type dataset --local-dir ./echopolicy-eval \
  --include "videos/**" "metadata/episodes.*" "metadata/video-files.json"

After a full download, verify the files and extract an episode for review:

cd echopolicy-eval
python3 scripts/verify_package.py
# Also decompress all archives and verify individual member checksums:
python3 scripts/verify_package.py --deep
python3 scripts/extract_episode.py 'close:efb6c8f2aa03d21d:0000' --output ./episode-review

The extraction command produces both original camera videos, the episode metadata and its batch HDF5. The relevant HDF5 group is printed by the command; other episodes in the same original batch remain in that file.

Reproducibility and limitations

The release provides evaluation recordings, logs, source hashes and dependency/runtime records. Policy source code, model weights and simulator asset bundles are not included in this dataset. Hosted VLM nondeterminism and service updates limit exact reproducibility.

The full benchmark includes episodes previously used during development; the results represent full-benchmark performance rather than an entirely held-out evaluation.

One Open episode (open:8745b89f28c7199a:0525) requested a legacy Push executor twice. Both initializations failed and produced no Push actions. Its evaluated outcome is included; details are in metadata/legacy-push-audit.json.

Attribution and rights

Benchmark and environment: Allen Institute for AI, MolmoSpaces. Evaluation policy: EchoPolicy-0, MagicLab. This dataset contains generated evaluation recordings and logs. Rights to the upstream benchmark, simulator assets and models remain with their respective owners.

For questions about the evaluation, use the official submission issue.