ddffwyb/EchoPolicy-0-MolmoSpaces-eval

EchoPolicy-0: MolmoSpaces Evaluation

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

• Updated benchmark entry: allenai/molmospaces#214
• Original benchmark entry: allenai/molmospaces#199
• Evaluation version: guoke3-v25-pick-phase-verification, full formal run.
• Release: v25-eval-20261011.
• 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 MP4 files in the fixed v25-eval-20261011 repository release. All recordings use H.264 with YUV420P at 624 × 352 pixels.

Results

Task	Episodes	Successes	At-end	Oracle	Official 95% interval
Close-v1	915	835	91.26%	91.26%	89.25–92.92%
Pick-v1.1	1000	847	84.70%	84.70%	82.33–86.80%
Open-v1	1000	702	70.20%	70.20%	67.29–72.95%
Pick & Place-v1	1000	556	55.60%	55.60%	52.50–58.65%

The mean success rate, with equal weight assigned to each of the four tasks, is 75.4392%. 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.

This update improves observed panel-contact geometry for Close, separates Pick/Place planning rejections from executable attempts, and strengthens Pick phase verification and the object-holding handoff to Place. All four tasks were reevaluated from the beginning with the frozen v25 implementation. This release uses the pretrained Pi_05 DROID checkpoint, with its file hashes recorded in the reproducibility metadata; no benchmark-specific model training was added.

• 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 fresh first 8 steps, then predict again from a new observation. Candidate inference uses microbatches of four. Geometric trajectories execute in chunks of at most 8 steps.
• Verification and retries: routine VLM checks every 16 executed environment steps, plus checks at tool completion or failure. Tools allow up to three execution attempts per executor/subgoal. Pick/Place planning has a separate limit of six initialization calls; a rejection before motion does not consume an execution attempt. Uncertain verification does not count as success and returns control to VLA after at most three checks.
• VLM: gemini-3.8-flash, thinking level medium, without an explicit output-token cap or global/batch VLM concurrency cap. Actual requests, responses and available provider model versions are indexed in metadata/vlm-request-coverage.json.
• 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). VLM waiting contributes to wall-clock time but does not advance simulation steps.
• Deployment: four A100 40 GB GPUs, each hosting one Pi_05/Echo service and sixteen simulation workers. Workers use their local policy service. Pi_05 memory preallocation is 0.35; model execution is serialized per service.
• Completeness: 3915 unique accepted episodes and 904,046 finite 7+1 actions. Two infrastructure interruptions were reviewed and retried; the first valid completed result for each episode was retained. All completed failures are included.

The four official CSVs were generated with the frozen official eval_to_csv.py, policy name EchoPolicy-0, and --success-condition both --dt 0.066. CSVs, export logs and the identical issue attachment ZIP 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/                         Four official CSVs, CSV ZIP and export logs
archives/                        Trajectories, evaluator logs, Agent logs and VLM traces
metadata/
  episodes.json                  Episode-to-video, trajectory and Agent-log mappings
  episodes.csv                   Flat episode index
  video-files.json               MP4 paths, sizes, hashes and full-decode checks
  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
  vlm-request-coverage.json       Episode coverage and actual generation settings
  official-export-audit.json     CSV hashes and official exporter identity
  reproducibility/               Frozen source, dependency, runtime and audit 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. Each selected video appears once, under videos/.

Trajectory archives preserve original HDF5 files and house/batch organization. These evaluation outputs include simulator state and evaluator labels. Evaluator-log archives contain episode sidecars, batch manifests, logs and attempt audits. Agent archives contain events and geometric-tool sensor inputs; VLM trace archives contain prompts, image payloads, generation settings and responses. Reviewed infrastructure interruptions and their replacement attempts remain traceable in the metadata and log archives.

All 3915 formal episodes have saved VLM traces. Use metadata/vlm-request-index.jsonl to locate requests in their archive, and metadata/vlm-request-coverage.json for episode coverage and actual provider settings.

Download and inspect

Install the Hugging Face CLI:

pip install -U huggingface_hub

Download this fixed release:

hf download ddffwyb/EchoPolicy-0-MolmoSpaces-eval \
  --repo-type dataset --revision v25-eval-20261011 \
  --local-dir ./echopolicy-eval

Download only the videos, episode table and video indexes:

hf download ddffwyb/EchoPolicy-0-MolmoSpaces-eval \
  --repo-type dataset --revision v25-eval-20261011 \
  --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, checkpoint and 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; these results represent full-benchmark performance rather than an entirely held-out evaluation.

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 this release, use the updated benchmark entry.