Bigenlight/fmrl-real-bc-rollouts

fmrl real-robot Flow BC (K=1) rollouts -- first50 + 15 Hz

Real UR7e + Robotiq 2F-85 GELLO eval episodes of the Flow BC K=1 baseline (actor = the frozen bc_flow module of an fmrl checkpoint, sampled once with Euler-10, no critic / candidate selection). Recorded by the same FM_RL real-eval stack (START -> HOLD GUI recorder) used for the companion GR00T N1.x VLA rollout dataset. private above is repo-visibility intent (all rows/files are private to the owner account, not a machine-checkable dataset property).

This dataset is scoped to fmrl Flow BC K=1, first50 + 15 Hz basis only (56D action, 8-knot H8 chunks, leader publish_rate_hz 15.0 -- the project's current shared evaluation standard, per FIRST50_15HZ_STANDARD_2026-09-22.md). The companion dataset Bigenlight/gr00t-n17-bigenlight-real-rollouts holds GR00T N1.x rollouts (BC/SVF/BoN adapter, 16-row action chunks, absolute 7D joint targets, 30 Hz ROS tick with optional action_repeat). The two use different robots/action spaces/execution stacks; do not merge success rates across them.

Scope: 15 Hz only

This repo publishes only the first50-15hz basis (real-flowbc-first50-15hz-{bowl, bowl_triple,cube}; the carrot variant exists as a launcher but recorded 0 episodes, see below). A separate, larger set of 30 Hz legacy Flow-BC rollouts (bc_from_svf_*, svfbase_full56_h8_bc* -- frozen bc_flow read out of full SVF/IFQL checkpoints, 30 Hz leader publish rate, contiguous H8) exists locally (~/gello_software_jazzy/ros2_ur_ws/log/eval/) but is deliberately not published here -- the project's current basis is first50 + 15 Hz and the two timing conditions (different chunk resolution, different publish rate, different BC checkpoint provenance) must never be pooled. If a future paper revision needs the 30 Hz family too, publish it as a clearly separate group set (or a separate dataset) rather than folding it in here.

basis	policy tick	leader publish rate	chunk	groups (published here)
first50-15hz	30 Hz	15 Hz (queue8/stride1/hold0, H8 knots)	8 knots / 0.533 s	real-flowbc-first50-15hz-{bowl,bowl_triple,cube}

This timing condition is NOT the same as GR00T's action_repeat mechanism in the companion VLA dataset (see the stream-layout note below) -- do not infer the policy clock from measured rate alone when comparing the two datasets.

group	success	partial	failure	unknown	unlabeled	invalid rec.	not trial	no scene ref	episodes	tasks	size
real-flowbc-first50-15hz-bowl	3	0	3	0	1	0	1	1	7	bowl	1.4 GB
real-flowbc-first50-15hz-bowl_triple	0	0	6	0	0	0	0	0	6	bowl_triple	2.2 GB
real-flowbc-first50-15hz-cube	0	0	9	4	0	0	0	0	13	cube	3.5 GB

Total: 26 episodes (3 success, 0 partial_success, 18 failure, 4 unknown, 1 unlabeled).

The carrot 15 Hz first50 launcher (bc_carrot_first50_h8_15hz_bc0) exists but recorded zero accepted episodes and is not in the table above -- do not read its absence as "0% success"; there was no trial. See FM_RL/BC_ROLLOUT_INVENTORY_2026-09-22.md section 1 (that audit also covers the unpublished 30 Hz legacy family, out of scope here).

Recording integrity

Two boolean fields, both per episode:

• recording_valid = episode.json.status == "ready" and not dropped -- the recording itself is intact (no dropped frames, no recorder-queue overflow). True for 26 of 26 episodes in this dataset, but the field must still be checked by any downstream consumer since other campaigns can produce status: "invalid" / termination_reason: "recording_error" recordings under recorder/GPU contention.
• is_trial = recording_valid and counts is non-empty and termination_reason != "start_rejected" -- an actual rollout happened. 1 episode(s) in this dataset pass the literal recording_valid check (status: "ready", dropped: 0) but have termination_reason: "start_rejected" and counts: {} -- the START was refused and no robot motion or policy tick was ever recorded. These are not trials and must be excluded from any denominator regardless of their outcome label.

**The success-rate denominator needs recording_valid AND is_trial AND a real outcome label** (i.e. exclude unknown/unlabeledtoo, see Labels below). The "not trial" and "invalid rec." columns above count episodes failingis_trial/recording_valid` respectively per group.

Labels

The folder suffix is the operator label and may carry a free-text note (__failure_59_cube -> outcome failure, note 59_cube). partial is normalised to partial_success. Folders suffixed __unlabeled, or any suffix that does not parse as a known outcome token, become outcome: "unlabeled" (never skipped -- unlike the upstream VLA uploader, every episode is uploaded).

Where the GUI record in outcome.json disagrees with the folder (hand-renamed folders), label_mismatch: true is set and both values are kept; the folder value is the row outcome. Mismatches fall into three categories (shared with the VLA audit): (a) the folder carries a real outcome but outcome.json.outcome is still unlabeled (GUI pass never actually ran on it); (b) folder and GUI both carry real outcomes but disagree (e.g. re-labeled after the folder was named); (c) the legacy success_0.5 string is outcome.json's alias for partial_success -- this is the same value under two names, not a real disagreement, and is flagged separately as label_mismatch_is_alias_only: true rather than label_mismatch: true. In this 15 Hz-only dataset: 0 category-(a)/(b) real mismatch(es) and 0 category-(c) alias-only non-mismatch(es) (none observed). The companion FM_RL/BC_ROLLOUT_INVENTORY_2026-09-22.md audit (section 2.3) covers the full local inventory including the unpublished 30 Hz family, where the real mismatches and all alias-only cases actually occur -- none of this dataset's 15 Hz episodes are affected.

unknown and unlabeled episodes must be excluded from success-rate denominators -- they are not failures, they are un-adjudicated. Label counts across all 26 episodes in this dataset: 3 success, 0 partial_success, 18 failure, 4 unknown, 1 unlabeled (21 of 26 are outcome-eligible for a success-rate computation once is_trial/recording_valid are also applied).

Per-episode files

Every file present in the source episode directory is uploaded -- nothing is excluded (unlike the VLA uploader, which drops the GUI-render diagnostic.mp4). Typical set: recording.h5 (see below), cam1.mp4 (SCENE), cam2.mp4 (WRIST), diagnostic.mp4 (post-hoc render, when present), episode.json, outcome.json, placement_reference.json (GUI dataset-overlay snapshot at START), summary.json, video_frames.json, render_status.json. Each episode folder also gets a tags.json with the same row the group's index.jsonl carries for that episode (minus the local source_path / files hash list, which stays in index.jsonl only to avoid duplicating large hash lists per file).

recording.h5 stream layout

HDF5 groups, each holding a payload (object array, one entry per tick) and a parallel timestamp_ns (int64) array; root attr episode_id. payload entries are uint8 byte buffers: JSON text for every stream except cam1/cam2 (raw JPEG bytes) and native_cam1/native_cam2 (raw camera-native frame bytes, metadata JSON alongside in native_cam{1,2}_meta).

group	rate/episode	payload content
robot_state	~500 Hz (UR RTDE feedback)	{name, position, velocity, effort, source_stamp_ns} (6 UR joints)
request	policy tick (30 Hz)	{request_id, control:{cmd, state(7=6 joints+gripper)}, cam{1,2}_source_ns, live_q, observation_arrival_monotonic_s} sent to the policy server
reply	policy tick (30 Hz)	{request_id, duration_ns, control:{ok, action(7)}} policy server response
leader_command	policy tick (30 Hz)	{request_id, position(6), policy_action(7), clamp_delta(6)} command sent to the GELLO leader / robot after safety clamp
gripper_command	policy tick (30 Hz)	{request_id, value}
gripper_state	~sparse	{position}
controller_command	~250 Hz	{data(7)} low-level controller setpoints
cam1 / cam2	policy tick	raw JPEG bytes, SCENE / WRIST
native_cam1 / native_cam2	camera-native rate	raw native-resolution frame bytes
native_cam1_meta / native_cam2_meta	camera-native rate	{source_stamp_ns}
lifecycle	1 per episode	{event: "accepted_start", accepted_ns}

The first50-15hz vs 30hz-legacy basis is corroborated by the measured leader_command/cam1 cadence (15.0 Hz vs 30.0 Hz exactly, per the inventory's stream audit) in this dataset only. Determine the policy clock from the tags (policy_hz, model_fps, publish_rate_hz), not from the measured stream rate, when comparing against the companion VLA dataset: there, a 15 Hz GR00T policy still logs leader_command at a full 30 Hz because each chunk row is repeated for 2 ticks (action_repeat: 2), so the measured rate alone does not tell you the policy tick rate across datasets -- only within this one.

Scene / placement reference

placement_reference.json is a GUI-only, display-side snapshot: which training take/episode frame was overlaid on the live camera feed at START for the operator's reference (task, dataset, split, selected_take, manifest, source-fingerprint hashes). It never reaches the policy input or the recorded frames (affects_policy_input: false) -- it does not affect what the robot executed.

1 of 26 episode(s) in this dataset have no placement_reference.json (has_placement_reference: false), so scene-matched comparisons must skip them -- it is the start_rejected episode in real-flowbc-first50-15hz-bowl (no run happened, see Recording integrity). The unpublished 30 Hz legacy family has more such gaps (including a svfbase_full56_h8_bc profile that predates the overlay feature entirely, and one episode whose overlay snapshot pointed at the wrong task's dataset -- display-only, did not affect execution); see FM_RL/BC_ROLLOUT_INVENTORY_2026-09-22.md section 3 for that full local inventory (out of scope for this repo).

Caveats

• This README's counts/labels/integrity claims are drawn from the concurrent read-only audit FM_RL/BC_ROLLOUT_INVENTORY_2026-09-22.md (repo: jun981015/FM_RL); consult it for the full per-episode breakdown behind every summary number here.
• robot_validated is not a field here -- every row in this dataset is a real hardware rollout log, but outcome labels are operator judgement calls made in the eval GUI from recorded video after the fact (labeler: "camera_viewer_pending" on every history entry in this scope -- provisional, not a formally audited label pipeline). Treat success / partial_success / failure as provisional until cross-checked.
• label_mismatch: true rows (folder suffix disagrees with outcome.json.outcome, e.g. a hand-renamed folder, or a folder written at an intermediate label and never renamed after a later re-label) are flagged per-row; check index.jsonl for the exact set before aggregating -- both label sources are kept so nothing is silently dropped. label_mismatch_is_alias_only: true rows are NOT real disagreements (see Labels).
• publish_rate_hz / model_fps are derived from group basis, not read from a per-episode live config dump; this matches the pinned safety-profile YAML sha for each launcher and is corroborated by the measured leader_command/cam1 cadence in the inventory's stream audit (15.0 Hz vs 30.0 Hz exactly, per profile) -- but it is a documented assumption in this script, not a per-row measurement of recording.h5.
• base_checkpoint is the frozen bc_flow-source checkpoint path referenced in the episode's metadata.files. For every group published in this dataset it is a params_0.pkl -- a natively-trained BC0 checkpoint (first50-15hz basis), not a submodule extracted from a later SVF/IFQL checkpoint. (The unpublished 30 Hz legacy family uses the other provenance -- a later-step params_<N>.pkl full SVF/IFQL checkpoint served with sampler: "bc" so only its frozen bc_flow submodule runs -- see the inventory doc section 1 if that family is ever published separately.)

Last update: 2026-09-22 21:03 KST