CLI¶
The complete index of qvla commands, including the tools a qvla group forwards to, generated
from their argument parsers. Each entry gives the command's one-line help and its usage; run a
command with --help for what each option means. A forwarded tool's usage line names its module
(python -m axis_qvla.realrobot.render source ...); qvla realrobot render source ... runs the same
command. To choose a command for a task, start from the tutorials.
qvla bench¶
the held-out benchmark: check the declaration, evaluate a checkpoint, read scores
qvla bench check¶
print the benchmark identity; exit 1 on any inconsistency
qvla bench check [-h] [--benchmark BENCHMARK] [--pretraining-prompts FILE]
qvla bench versions¶
list the benchmark versions under conf/benchmarks/ and their hashes
qvla bench versions [-h]
qvla bench eval¶
evaluate ONE checkpoint on the benchmark (the campaign driver, fully composed)
qvla bench eval [-h] [--checkpoint CHECKPOINT] [--checkpoint-id ID] [--from-run RUN_ID] [--verify-checkpoint] [--config CONFIG] [--policy HOST:PORT] [--run RUN] [--serving {bare,conditioned}] [--asset-dir ASSET_DIR] [--render-root RENDER_ROOT] [--task-data-root TASK_DATA_ROOT] [--guidance-scale GUIDANCE_SCALE] [--backend {openpi,smolvla,xvla}] [--benchmark BENCHMARK] [--host HOST] [--gpus GPUS] [--workers-per-gpu WORKERS_PER_GPU] [--batch BATCH] [--mem-fraction MEM_FRACTION] [--dry-run] [--tally-only] [--seed-set {test,dev}] [--off-protocol] [--ft-seed SEED] [--pretrain-seed SEED] [--print]
qvla bench score¶
score results files (all-24 and per-task, with seed-bootstrap intervals), or list the run summaries under a campaign root
qvla bench score [-h] [--run-root RUN_ROOT] [--run RUN] [--benchmark BENCHMARK] [--view {all24,informative_20,usable_12}] [--bootstrap B] [--per-task] [--allow-off-protocol] [files ...]
qvla bench check-server¶
check that a running policy server speaks the protocol eval uses (docs/reference/policy-server.md)
qvla bench check-server [-h] --policy POLICY [--benchmark BENCHMARK] [--task TASK] [--calls CALLS] [--timeout TIMEOUT] [--report REPORT] [--host HOST]
qvla data¶
get and verify the benchmark's data against the release checksum manifest
qvla data verify¶
compare the checksum manifest with the Hub (no download) or a local tree
qvla data verify [-h] [--checksums FILE] [--local LOCAL] [--repo REPO]
qvla data get¶
download a published data set, verified file by file against the release checksum manifest
qvla data get [-h] [--include GLOB] [--with-init] [--serving {pi05,smolvla,xvla}] [--dest DEST] [--host HOST] [--checksums FILE] [--results-map FILE] target
qvla init¶
write ./qvla.toml: where this machine keeps the data, runs and environments
qvla init [-h] [--root DIR] [--force] [--gpus N]
qvla paper¶
reproduce the paper's results -- every figure value, the main results and the in-text numbers -- from the published per-seed files
qvla paper tables¶
rebuild every figure value, the main results and every in-text number; --check compares with the pack's expected/
qvla paper tables [-h] [--results RESULTS] [--out OUT] [--check] [--cache CACHE] [--results-map FILE] [--checksums FILE] [--benchmark BENCHMARK]
qvla paper label-study¶
the quality-label distributions and their rank correlation with human judgments, from published files
qvla paper label-study [-h] --pretraining-data DIR [--study STUDY] [--expected EXPECTED] [--out OUT] [--check]
qvla paper reeval¶
re-evaluate released checkpoints -- one result, the main results or a figure -- and write the paper-layout tables from your own runs
qvla paper reeval [-h] (--result ID | --table NAME | --figure NAME | --list) [--backbone {pi05,smolvla,xvla}] [--out OUT] [--host HOST] [--print] [--collect]
qvla labels¶
compile, export and check the paper's quality labels and training weights
qvla labels build-paper¶
compile every paper run's labels from the published pretraining data
qvla labels build-paper [-h] --pretraining-data DIR --episodes DIR --out OUT [--only ONLY]
qvla labels export¶
write compiled labels' legacy files, byte-equal to what the paper's runs read
qvla labels export [-h] --format {openpi-legacy,lerobot-legacy} --out OUT COMPILED
qvla labels check¶
verify a compiled-labels tree against the paper's declaration and re-export every file the checksum manifest lists from it, byte for byte
qvla labels check [-h] COMPILED
qvla results¶
build qvla.results/v1 files from an evaluation, and score, contrast and re-check them under the reporting rules
qvla results score¶
score results files under the reporting rules: all-24 and per task, with seed-bootstrap intervals (as bench score <files>)
qvla results score [-h] [--benchmark BENCHMARK] [--view {all24,informative_20,usable_12}] [--bootstrap B] [--per-task] [--allow-off-protocol] files [files ...]
qvla results contrast¶
contrast two methods' results files, one per training seed on each side: 100 (SR_A - SR_B) / SR_BC per replicate with --relative, else percentage points
qvla results contrast [-h] [--benchmark BENCHMARK] --a A [A ...] --b B [B ...] [--bc BC [BC ...]] [--relative] [--allow-off-protocol] [--results RESULTS] [--cache CACHE] [--host HOST]
qvla results reeval¶
is a re-read of a checkpoint consistent with its published read? (a reproducibility check; exit 1 if not)
qvla results reeval [-h] --published PUBLISHED --new NEW [--view {all24,informative_20,usable_12}] [--allow-different-checkpoints] [--benchmark BENCHMARK]
qvla results build¶
turn one qvla bench eval run into a qvla.results/v1 file
qvla results build [-h] --run-root RUN_ROOT --run RUN --ft-seed FT_SEED --pretrain-seed PRETRAIN_SEED [--checkpoint-ref CHECKPOINT_REF] [--log-root LOG_ROOT] --out OUT [--benchmark BENCHMARK]
qvla realrobot¶
real-robot steps (workspace … serve) and tools: kinematic renderer, REALALIGN labels, FR3 data and configs
qvla realrobot render¶
python -m axis_qvla.realrobot.render
qvla realrobot render source¶
a source_episode.json from a recorded trace
python -m axis_qvla.realrobot.render source [-h] --trace TRACE --scene-dir SCENE_DIR --scene-xml SCENE_XML --task-id TASK_ID --attempt-id ATTEMPT_ID --prompt PROMPT --out OUT
qvla realrobot render prepare¶
the render input tree of the simulation demonstrations in qvla-v1-realworld
python -m axis_qvla.realrobot.render prepare [-h] [--data DATA] [--out OUT] [--tasks TASKS [TASKS ...]] [--host HOST]
qvla realrobot render render¶
render source episodes to LeRobot v3 tasks
python -m axis_qvla.realrobot.render render [-h] [--episodes EPISODES [EPISODES ...]] [--data DATA] [--tasks TASKS [TASKS ...]] [--robots ROBOTS] [--gripper-xml GRIPPER_XML] [--unlocked-gripper] [--profile PROFILE] --out OUT
qvla realrobot render droid8d¶
the DROID-8D training view of a render
python -m axis_qvla.realrobot.render droid8d [-h] --src SRC --out OUT
qvla realrobot render labels¶
attempt-keyed phase curves -> the pretraining data's curve store
python -m axis_qvla.realrobot.render labels [-h] --pretraining-data CORPUS [--attempt-curves ATTEMPT_CURVES] [--export-curves EXPORT_CURVES] [--alignment-map ALIGNMENT_MAP] --out OUT
qvla realrobot render gates¶
the hard gates -> render_gates.json
python -m axis_qvla.realrobot.render gates [-h] --out OUT [--droid8d DROID8D] [--curves CURVES] [--attempt-curves ATTEMPT_CURVES]
qvla realrobot render report¶
reported metrics against a reference render of the same attempt
python -m axis_qvla.realrobot.render report [-h] --task TASK --episode EPISODE --ref-task REF_TASK --ref-episode REF_EPISODE --out OUT
qvla realrobot render overlay¶
the outlines of the rendered table and robot on a real camera frame
python -m axis_qvla.realrobot.render overlay [-h] [--profile PROFILE] --image IMAGE --scene SCENE [--robots ROBOTS] --out OUT [--joints JOINTS JOINTS JOINTS JOINTS JOINTS JOINTS JOINTS]
qvla realrobot realign¶
python -m axis_qvla.realrobot.realign
qvla realrobot realign build¶
phase curves -> gated signal artifacts
python -m axis_qvla.realrobot.realign build [-h] --out OUT [--export-curves EXPORT_CURVES] [--alignment-map ALIGNMENT_MAP] [--prompts PROMPTS] [--roots ROOTS] [--curves-store CURVES_STORE] [--min-fit-pool MIN_FIT_POOL] [--valid-from-data] [--runs {bc,conditioning,awr,filtering} [{bc,conditioning,awr,filtering} ...]]
qvla realrobot realign gates¶
re-run the gates over a built directory
python -m axis_qvla.realrobot.realign gates [-h] --out OUT [--runs {bc,conditioning,awr,filtering} [{bc,conditioning,awr,filtering} ...]]
qvla realrobot realign compile¶
compile the four stage-1 runs (bc, conditioning, awr, filtering) to compiled labels
python -m axis_qvla.realrobot.realign compile [-h] --built BUILT --out OUT [--runs {bc,conditioning,awr,filtering} [{bc,conditioning,awr,filtering} ...]]
qvla realrobot realign phase¶
phase curves of rendered episodes from the published recordings
python -m axis_qvla.realrobot.realign phase [-h] --rendered RENDERED --data DATA --out OUT [--tasks TASKS [TASKS ...]]
qvla realrobot realdata¶
python -m axis_qvla.realrobot.realdata
qvla realrobot realdata idle-ranges¶
the idle filter: kept segments, optionally the exported dataset
python -m axis_qvla.realrobot.realdata idle-ranges [-h] --dataset DATASET --out OUT [--export EXPORT] [--grip-guard GRIP_GUARD] [--grip-delta GRIP_DELTA]
qvla realrobot realdata tail¶
tail-10 chunk-start ranges (openpi filter_dict_path file)
python -m axis_qvla.realrobot.realdata tail [-h] --dataset DATASET --out OUT [--n N]
qvla realrobot realdata crop224¶
rebuild a letterboxed dataset's images as centre-crop-224
python -m axis_qvla.realrobot.realdata crop224 [-h] --dataset DATASET --recordings RECORDINGS --out OUT [--window WINDOW] [--episodes EPISODES [EPISODES ...]]
qvla realrobot realdata training-set¶
a DROID-layout recording in the per-task layout the π0.5 loader reads
python -m axis_qvla.realrobot.realdata training-set [-h] --dataset DATASET --out OUT [--tail TAIL]
qvla realrobot fr3¶
python -m axis_qvla.realrobot.fr3
qvla realrobot fr3 check¶
the line record, its ArmSpecs, presets, a server's handshake, a norm file
python -m axis_qvla.realrobot.fr3 check [-h] [--line LINE] [--runs RUNS] [--presets] [--run RUN] [--handshake HANDSHAKE] [--dataset DATASET] [--norm NORM] [--embodiment {fr3,realalign_sim,fr3_droid}]
qvla realrobot fr3 norm¶
norm statistics over a training set, in the model's state order
python -m axis_qvla.realrobot.fr3 norm [-h] --dataset DATASET --embodiment {fr3,realalign_sim,fr3_droid} [--out OUT] [--host HOST]
qvla realrobot fr3 local-runs¶
run specs that name this machine's real-world inputs by content
python -m axis_qvla.realrobot.fr3 local-runs [-h] [--host HOST] [--line LINE] [--runs RUNS [RUNS ...]]
qvla realrobot fr3 serve-config¶
the openpi config a policy server of a real-world run is started with
python -m axis_qvla.realrobot.fr3 serve-config [-h] --run RUN [--out OUT] [--host HOST]
qvla train¶
train a run spec in its backbone's environment (--print: the command only)
qvla train [-h] [--host HOST] [--smoke] [--no-bind] [--fsdp-devices N] [--resume] [--print | --render] run_spec
qvla train smoke-fixture¶
build the 2-task x 3-episode fixture
qvla train smoke-fixture [-h] [--host HOST] --episodes DIR
qvla train merge-check¶
check that a plain merge accepts these task directories
qvla train merge-check [-h] tasks [tasks ...]
qvla train pretrain¶
pretrain a backbone on your compiled labels with the paper's budget
qvla train pretrain [-h] --labels NAME --backbone {pi05,smolvla,xvla} [--run NAME] [--seed N] [--host HOST] [--fsdp-devices N] [--print]
qvla train finetune¶
fine-tune a pretraining run with the paper's recipe on the 24 held-out tasks
qvla train finetune [-h] --from RUN_ID [--cond] [--run NAME] [--host HOST] [--fsdp-devices N] [--print]
qvla env¶
build and check the training and evaluation environments
qvla env build¶
build an environment from its pinned lock, then check it
qvla env build [-h] [--print] [--rebuild] environment
qvla env check¶
check a built environment against its pins
qvla env check [-h] [--python PYTHON] [--checkout CHECKOUT] environment