LFF Bench Agent Replays — v1.5
LFF Bench Agent Replays contains dense robot-control trajectories reconstructed from completed agent evaluations. It includes successful and failed manipulation attempts, retries, task-event annotations and available agent narration. The robot is a simulated Franka Panda in MuJoCo / robosuite. This release consolidates 41 previously published packages into one dataset with globally consistent indices and traceable label revisions.
中文概览： 本版包含 2,196 条轨迹、3,402,853 个控制帧、47.26 小时仿真操作，覆盖 34 个任务。成功与失败轨迹都保留；标签来自任务判定器、执行日志及有来源哈希验证的既有证据复核，并非人工逐帧标注。
Release contents
| Quantity | Count |
|---|---|
| Original evaluation runs | 1,635 |
| Episodes / trajectories | 2,196 |
| Control frames at 20 Hz | 3,402,853 |
| Simulated execution hours | 47.26185 |
| Successful terminal episodes | 673 |
| Failed terminal episodes | 1,523 |
| Attempts inside episodes | 2,283 |
| Successful / failed attempts | 673 / 1,610 |
| Tasks | 34 |
| Cameras | 2 × RGB, 256 × 256, 20 fps |
Hours count physical control time once, not once per camera, and exclude agent reasoning, network latency and conversion time. Each cross-protocol attempt becomes one episode; within-protocol episodes can contain several attempts without resetting the scene. Episode success refers to its last attempt. A successful episode can contain earlier failures and recovery actions.
This is the original 34-task collection plus its completed 2026-11-02 backfill. The separately running new compound-task evaluations are outside this frozen release. Source batches include formal evaluations, controlled ablations, pilot/smoke and diagnostic runs. Their model, harness, seed, history, feedback and simulator provenance remain available. The collection's success ratio is not a controlled model leaderboard score.
| Model name in source logs | Episodes |
|---|---|
| gpt-6-astra | 346 |
| claude-sonnet-5-5 | 519 |
| claude-opus-5-5 | 470 |
| gpt-6.1-sol | 172 |
| claude-fable-5-1 | 46 |
| deepseek-flash | 136 |
| deepseek-v4.1-flash | 507 |
Example trajectory
Episode 915 shows a real tray-delivery attempt: bottles are loaded, red/blue bottles topple during transport, and the agent restores them before successful delivery. This episode has 3,322 control frames (166.10 simulated seconds), 52 command records, three regressions and three restorations. It is attempt 2 of a run with history FS; its terminal success label does not mean it was error-free throughout. The images are decoded from the published source video, not generated illustrations. Full annotation.
Format and loading
The dataset follows LeRobotDataset v3.0: numeric Parquet tables, multi-episode MP4 files, relational episode metadata and normalization statistics. Additional LFF annotations live in lff/. Video files are copied byte-for-byte from accepted source packages, without another encoding pass; existing AV1/yuv420p compression remains lossy.
from pathlib import Path
from lerobot.datasets.lerobot_dataset import LeRobotDataset
# Local checkout or downloaded dataset directory:
dataset = LeRobotDataset(
'OWNER/DATASET',
root=Path('/path/to/lffbench-v1.5'),
)
frame = dataset[0]
print(frame['observation.state'], frame['action'], frame['task'])
Use a LeRobot installation supporting v3.0 and an AV1-capable video backend. The release was tested with the local LeRobot reader; the exact environment is recorded in lff/release/environment.json. Older LeRobot 0.1 readers are not compatible with this layout. The repository name above is supplied by the eventual uploader; local loading does not require an upload.
meta/info.json schema, FPS, total counts, path templates
meta/tasks.parquet global task-instruction index
meta/episodes/ global episode boundaries and video offsets
meta/stats.json numeric statistics; recomputed sampled image stats
data/chunk-*/file-*.parquet dense numeric controls, observations and annotations
videos/<camera>/chunk-*/*.mp4 original AV1 camera streams
lff/annotations/ detailed per-episode JSON
lff/episodes.jsonl flat episode index and outcomes
lff/runs.jsonl original runs and cross-attempt links
lff/tasks.json taxonomy and distinct instance variants
lff/sim_state/ physical simulator state per control frame
lff/final/ state and images after the last control action
lff/release/ provenance, label audit, validation, split groups
lff/code/ source-package replay/export utilities
SHA256SUMS complete release file checksums
One MP4 is not necessarily one trajectory. Read the episode metadata to locate its camera file and timestamps. lff/release/episode_mapping.jsonl maps every global episode back to its source package, local episode, run and attempts. episode_uid is stable across index changes. The old cumulative versions and their already included supplements must not be counted again alongside v1.5.
The Parquet preview configuration exposes numeric training rows; use LeRobot for synchronized video access. There are 34 logical tasks and 37 distinct instruction labels in meta/tasks.parquet. For 31 count/sequence episodes, generic N or “given per instance” templates are replaced with the recorded target number or order. 27 were also checked against the exact original first-message task text; the others are explicitly marked as resolved from preserved instance parameters using the frozen task formatter. Original templates and per-episode evidence remain in instruction_audit.jsonl.
Observation and action semantics
| Field | Meaning |
|---|---|
observation.images.image |
External RGB camera |
observation.images.wrist_image |
Wrist RGB camera |
observation.state |
float32[8]: world end-effector xyz (m), axis-angle rotation (3), two finger joint positions |
observation.joint_pos, observation.joint_vel |
float32[7], arm joint radians and radians/s |
observation.eef_quat |
float32[4], xyzw quaternion |
observation.gripper_width |
float32[1], meters |
action |
float32[7], normalized OSC_POSE translation/rotation inputs and gripper command |
timestamp |
Episode-relative simulated seconds |
Each observation is captured before its row's action. The post-final-action state and images are separate under lff/final/. Position and rotation controller scales are 0.1 m and 0.5 rad; gripper −1 opens and +1 closes. These are normalized controller inputs, not absolute poses or meter-valued raw actions. The agent originally issued coarse commands such as move_eef; the skill controller expanded them into the dense action stream.
Original agents typically saw 512 × 512 JPEG observations at command boundaries. Dense 256 × 256 replay frames are not a record of images that the agent saw at every control step. No uniform depth, segmentation, object bounding-box, object 6D-pose or human per-frame caption annotation is provided in the standard table.
Annotation levels and label revisions
Frame annotations associate controls with an attempt and command. annotation.phase is 0 for movement, 1 for gripper motion and 2 for judge/settling frames. Phase 2 is not a new agent decision and its action need not be zero. annotation.attempt_success repeats the attempt's eventual result (1 / 0); it is not instantaneous goal attainment. annotation.task_event is an event count, not an event category.
Per-episode JSON contains task instructions and instance parameters; run/model/harness/condition provenance; attempt outcomes and feedback; commands and half-open [frame_start, frame_end) intervals; task-monitor events with episode frame indices; and available agent text, reasoning summaries and tool-call information. Task-internal event.step can differ from episode frame; use frame for video alignment.
Task causes now consistently use the frozen L1–L5 taxonomy in lff/release/evidence/causes.json. This corrects 452 episode-level cause values, including placeholders and inconsistent paper-table labels. Original values are retained. Historical task IDs and level_code values such as L6/L7 are preserved as identifiers and are not the canonical failure-cause labels. Task design classes and monitor-provided cause fields are not independently verified per-frame causal explanations.
Recovery classes describe the original run: CS = first-attempt success without an identified qualifying recovery; IR = recovery within the first attempt; XR = success on a later attempt after failure; F = all attempts failed. The revision uses 148 existing evidence-audit entries only after matching batch, manifest ID, task, seed and the hashes of their original event/trajectory files. Those are existing evidence reviews (their reviewers and reasoning are preserved), not a new claim of human annotation.
outcome_class_status explicitly distinguishes:
evidence_reviewed: an applicable, source-hash-verified existing review;rule_derived: a retained classification from exporter rules, not individually reviewed;unresolved: 49 runs lack sufficient recovery-type evidence; canonical class is null and the originalS?is preserved;not_applicable: the classification is not defined for that history/protocol condition.
These statuses accompany the class in detailed annotations, the flat episode index and the run index. Attempt success/failure is unchanged by recovery-type cleanup. Existing task-event and automatic error flags are preserved; they can differ from the stricter reviewed definition of physical failure followed by an active corrective action. See label_audit.jsonl for old/new labels and evidence.
Agent narration is optional and is not a complete private chain of thought or complete conversation archive. Empty text fields do not imply that an agent did not reason. The original exporter truncates shell command text to 2,000 characters and final reports to 6,000 characters. feedback_shown is the feedback actually presented; feedback_f2, hidden instance parameters and future success labels are privileged offline annotations and must not silently become online policy inputs.
Splits and intended use
The LeRobot train split names the entire collection. No official held-out train/validation/test partition is claimed. lff/release/split_groups.json groups all episodes by task and seed, across models, conditions and retries, so downstream users can make a documented partition without splitting closely related attempts. Keep shared-first-attempt and retry relatives together; inspect simulator/condition provenance when designing comparisons. Recompute normalization on the training subset after defining a held-out split; the packaged statistics cover the entire collection.
Potential uses include imitation learning on selected successful segments, recovery-policy training, failure detection, command/action alignment and studies of feedback or history. Success-only behavior cloning should still distinguish clean motion from failure/recovery and judge frames. This is a heterogeneous agent-generated simulation collection with few distinct instances per task; it is not a balanced set of human expert demonstrations and does not establish real-robot transfer.
MuJoCo NPY state supports reconstructing the physical configuration with matching scene/code. It is not a complete checkpoint of Python task monitors, controller internals or agent context. Replaying float32 actions alone can diverge under contact dynamics. The included exporter scripts are not a complete simulator dependency bundle.
Validation and provenance
Accepted source runs had already passed their source code/instance/report/outcome replay checks; rejected candidates are accounted for separately from failed task trajectories. The merge never changes task success labels, actions, observations, raw physics files or video bytes.
This release verifies all 3,402,853 numeric rows against their sources after the three necessary index remappings, all 8,913 copied binary files by SHA-256, metadata bounds, dense attempt labels, run/attempt links and per-camera frame counts. It also loads 160 actual frames through LeRobot across source packages, tasks and video files; decoded pixel difference from direct video decoding is 0. Every video is header/count checked and byte compared; the release does not claim a fresh full decode of every frame or a rerun of every simulator trajectory.
All scalar/vector numeric normalization statistics are recomputed on the merged collection, including remapped indices. Camera statistics are recomputed from three deterministic quarter/mid/three-quarter frames per episode (fewer if frame indices coincide), using all RGB pixels and float64 histogram moments. This fixes the old uint8-squared variance overflow that produced zero within-episode image standard deviations. Per-episode and global camera statistics are both updated. These are stratified sample estimates, not a full-frame census; sampled indices, old global statistics and the calculation method are recorded in image_stats.json. Complete checksums, environment versions, source manifests and the verification report accompany the release.
License and publication
Dataset license: pending owner selection. The local repository has no explicit dataset license; fill in the chosen license before public publication.
The intended Hub repository is pending owner selection. This release is prepared locally and has not been uploaded. Repository creation, visibility and the actual transfer are separate steps. See UPLOAD.md.
The README front matter follows the Hugging Face dataset-card format. Release-specific claims above are supported by the included local audit artifacts.
- Downloads last month
- 235