Commit Graph

  • fd18beb3a1 review: address CarolinePascal feedback Pepijn Kooijmans 2026-05-18 12:03:25 +02:00
  • 26cb38a7d0 feat(smolvla2): startup task picker, /vlm mode toggle, interactive VQA overlay Pepijn 2026-05-18 11:20:57 +02:00
  • 015c88cf0d Frame count is now derived from the upstream .npy length feat/robometer-rm Khalil Meftah 2026-05-18 10:57:16 +02:00
  • bfb8cfb432 fix(smolvla2): flatten say tool_calls into <say> marker before tokenizing Pepijn 2026-05-18 10:47:31 +02:00
  • 5e3b9ba82c tune(smolvla2): override optimizer_lr to 2.5e-5 for pretrained-LM fine-tuning Pepijn 2026-05-18 10:41:13 +02:00
  • 0164725af8 fix decord Khalil Meftah 2026-05-18 10:39:51 +02:00
  • 083d3cd419 tune(smolvla2): soften flow:text loss split from 10:1 to 5:1 Pepijn 2026-05-17 16:00:08 +02:00
  • 34274c6f70 scripts: add Robometer parity checks (upstream example videos + LIBERO) Khalil Meftah 2026-05-17 15:41:31 +02:00
  • f6a13b1338 Add Robometer reward model Khalil Meftah 2026-05-17 14:59:23 +02:00
  • bf996c7938 fix(datasets): render flow-only low_level recipes instead of dropping them Pepijn 2026-05-17 13:20:39 +02:00
  • 0d88eaf8eb test(smolvla2): attention masking of the language target span Pepijn 2026-05-16 18:28:44 +02:00
  • 3cd348ffe2 fix(smolvla2): causal mask on the text-CE target span (THE collapse bug) Pepijn 2026-05-16 18:24:44 +02:00
  • db03fc6dc4 fix(smolvla2): select_message must decode from the language position Pepijn 2026-05-16 15:05:16 +02:00
  • 56068d37ea fix(smolvla2): default load_vlm_weights=True — don't train from scratch Pepijn 2026-05-15 16:44:00 +02:00
  • e727688052 annotate: telegraphic subtasks — ≤4 words, verb+object, consistent nouns Pepijn 2026-05-15 14:14:42 +02:00
  • f1a0a663cc fix(inference): gibberish detector catches long repetition collapse Pepijn 2026-05-15 13:52:26 +02:00
  • 6e64c20cf1 runtime: stop seeding plan/memory from the dataset (unused) Pepijn 2026-05-15 13:47:33 +02:00
  • b29cccb37e runtime: restore the subtask hierarchy — generated subtask drives actions Pepijn 2026-05-15 13:43:04 +02:00
  • f161e27e96 recipe+runtime: condition the action expert on the task, not the subtask Pepijn 2026-05-15 13:40:15 +02:00
  • 01dcb4c292 fix(pi05): update pi05 with transformers v5.4.0 interface (#3603) Haoming Song 2026-05-15 17:37:05 +08:00
  • d5f293a1c9 recipe+runtime: VQA + subtask only — drop plan & memory Pepijn 2026-05-15 08:02:06 +02:00
  • bd9619dfc3 feat(encoding parameters): adding support for user provided video encoding parameters (#3455) Caroline Pascal 2026-05-14 23:46:42 +02:00
  • 0a4a7c40ad docs(cheat sheet): create cheat sheet (#3602) Nikodem Bartnik 2026-05-14 15:11:35 +02:00
  • ca9028ad64 docs(quickstart): adding rollout (#3598) Nikodem Bartnik 2026-05-14 12:32:39 +02:00
  • 95033733fc deps: add sentencepiece to the pi extra (FAST action tokenizer) Pepijn 2026-05-13 17:52:55 +02:00
  • c3503b774f fix(debug): dumper now shows real stream + target flags Pepijn 2026-05-13 16:43:51 +02:00
  • 99ebee4d16 annotate: tighter subtask + memory prompts (≤5 / ≤10 words) Pepijn 2026-05-13 16:28:09 +02:00
  • a8ca5128b8 fix(annotate): re-emit plan at every subtask boundary Pepijn 2026-05-13 16:26:49 +02:00
  • dd97c33814 refactor(annotate): plan = summary of still-todo subtasks, drop VLM call Pepijn 2026-05-13 15:55:02 +02:00
  • fa45ba631b fix(policies,recipe): register PI052Config + allow flow-only sub-recipes Pepijn 2026-05-13 15:51:47 +02:00
  • ffd8c92ce5 fix(inference): always emit Plan:/Memory: labels in the high-level prompt Pepijn 2026-05-13 15:42:29 +02:00
  • 841d3c47e1 feat(debug): LEROBOT_DUMP_RECIPE_SAMPLES=N dumps the first N rendered samples Pepijn 2026-05-13 15:21:46 +02:00
  • 2c920ab178 refactor(recipes): consolidate to shared hirobot.yaml + audit fixes Pepijn 2026-05-13 15:16:28 +02:00
  • 9f630e2a41 fix(recipes,training): stop tool prompt leak + drop subtask copy-supervision Pepijn 2026-05-13 14:59:01 +02:00
  • 7a32f8a72a refactor(recipes): π0.5-style split — action expert conditions on subtask only Pepijn 2026-05-13 14:13:07 +02:00
  • 129aa207e3 fix(smolvla2,pi052): training-correctness audit fixes Pepijn 2026-05-13 14:08:06 +02:00
  • e3ad1c59fc feat(recipes): add plan_generation sub-recipe to smolvla2 + pi052 blends Pepijn 2026-05-13 13:51:37 +02:00
  • 9ff62cb08c docs(recipes): trim header comments, drop diversity-knobs note in run_hf_job Pepijn 2026-05-13 12:55:03 +02:00
  • b2aa372fcf refactor(recipes): fold memory into action_execution, drop interjection, fuse smolvla2 forward Pepijn 2026-05-13 12:51:09 +02:00
  • 058b8f3958 refactor(recipes): two-flavor design — one fused action_execution + text-only events Pepijn 2026-05-13 12:35:51 +02:00
  • b873fe454c perf(pi052): full fusion — text + FAST + flow in ONE backbone forward Pepijn 2026-05-13 12:28:38 +02:00
  • 83d7250a22 fix(recipes): low_level_execution needs if_present:subtask guard too Pepijn 2026-05-13 12:22:45 +02:00
  • 35f9063a6c perf(pi052): fuse text + FAST loss into a single prefix forward Pepijn 2026-05-13 12:08:34 +02:00
  • 17c0800461 fix(pi052): FAST loss masking + predict_actions gating + smolvla2 review Pepijn 2026-05-13 12:05:37 +02:00
  • c8763e0ad5 fix(pi052): four real bugs in the modeling code + flip defaults Pepijn 2026-05-13 11:58:40 +02:00
  • 0f4faddc01 feat(pi052): auto-fit FAST tokenizer per-dataset before training Pepijn 2026-05-13 11:52:31 +02:00
  • 8dc0af3c28 feat(pi052): FAST action CE loss + knowledge insulation + processor wiring Pepijn 2026-05-13 11:46:21 +02:00
  • 9db9c35cb4 fix(config): add lora_alpha to PeftConfig (#3573) Cheng Yin 2026-05-13 17:09:19 +08:00
  • 8eba704f15 Revert "chore(training): align pi052_hirobot.slurm with the operator's actual command" Pepijn 2026-05-13 11:03:58 +02:00
  • ecbac17196 chore(training): align pi052_hirobot.slurm with the operator's actual command Pepijn 2026-05-13 11:03:09 +02:00
  • 12cce8f2cc fix(smolvla2): align flow_loss_weight default with Pi 0.5 paper's α=10 Pepijn 2026-05-13 11:02:17 +02:00
  • ef5879a02a feat(pi052): π0.5 v2 — full reproduction of the π0.5 paper recipe Pepijn 2026-05-13 10:59:26 +02:00
  • 1d24301b67 chore(training): STEPS=15000 default + dropout walked back to 0.30/0.30/0.20 Pepijn 2026-05-13 10:46:19 +02:00
  • fe96b28c74 Fix policy.path not working in YAML config files (#3145) Jash Shah 2026-05-13 00:45:27 -07:00
  • 3a20ea337e feat(smolvla2-runtime): --text_min_new_tokens / --text_temperature CLI debug knobs Pepijn 2026-05-12 21:39:33 +02:00
  • b6fb536460 chore(training): bump plan/memory dropout to 0.50 to force vision-grounding Pepijn 2026-05-12 21:30:51 +02:00
  • 2438df1307 chore(dependencies): update uv.lock (#3561) Steven Palma 2026-05-12 21:20:26 +02:00
  • f218d5ab30 feat(episodes): adding support for metadata based episodes filtering (#3530) Caroline Pascal 2026-05-12 20:44:11 +02:00
  • bfd3bb1791 fix(smolvla2): handle batched sample indices in chat tokenizer pepijn 2026-05-12 16:56:13 +00:00
  • 4908433f9a chore(training): align smolvla2_hirobot.slurm with what's actually run Pepijn 2026-05-12 18:45:38 +02:00
  • 6ce1f36002 fix(smolvla2): supervise high-level head with *current* subtask at every frame Pepijn 2026-05-12 18:42:59 +02:00
  • 731576be80 chore(smolvla2-runtime): auto-fire one tick at dry-run startup Pepijn 2026-05-12 18:34:42 +02:00
  • 47fb8318b1 chore(training): widen augmentation envelope after live-robot diagnostic Pepijn 2026-05-12 18:25:41 +02:00
  • 53172873e3 chore(smolvla2-runtime): probe obs once at dry-run startup Pepijn 2026-05-12 18:21:58 +02:00
  • fcdae0ce8e chore(smolvla2-runtime): tensor-level obs print for both inference paths Pepijn 2026-05-12 18:19:18 +02:00
  • 4852b9f952 feat(smolvla2-runtime): --dataset.augment_at_inference for the bisection test Pepijn 2026-05-12 18:14:57 +02:00
  • 0410705aff chore(smolvla2-runtime): print live state vector once at startup Pepijn 2026-05-12 18:12:27 +02:00
  • 398a8cf730 chore(smolvla2-runtime): log first-tick resize so train/inference match is verifiable Pepijn 2026-05-12 18:06:00 +02:00
  • ab5c1dc392 fix(smolvla2-runtime): match training visual distribution on robot frames Pepijn 2026-05-12 17:59:24 +02:00
  • 1292304c42 fix(smolvla2): suppress all special tokens during min_new_tokens window Pepijn 2026-05-12 17:49:53 +02:00
  • b95eebff77 fix(smolvla2): force min_new_tokens + sampling so memorised LM emits something Pepijn 2026-05-12 17:48:08 +02:00
  • fbcac95662 feat(smolvla2-runtime): scrollback in autonomous panel + empty-gen counter Pepijn 2026-05-12 17:42:13 +02:00
  • b9db4d21a2 fix(smolvla2): high-level steps must run before LowLevelForward refills Pepijn 2026-05-12 17:38:06 +02:00
  • aecb80a9d2 feat(smolvla2-runtime): overfit/memorisation diagnostics on the panel Pepijn 2026-05-12 17:31:04 +02:00
  • c98c695127 feat(smolvla2-runtime): 'rephrase:' prefix to swap task string in place Pepijn 2026-05-12 17:26:59 +02:00
  • d528078aca fix(smolvla2-runtime): allow task switching mid-run via 'task:' prefix Pepijn 2026-05-12 17:24:16 +02:00
  • a648da0455 fix(smolvla2): unblock action dispatch when high-level LLM stalls loop Pepijn 2026-05-12 17:23:09 +02:00
  • d866c2c9fd fix(smolvla2): only regenerate chunk when queue is fully drained Pepijn 2026-05-12 17:15:02 +02:00
  • 04125492e4 fix(datasets): expand torchcodec platform coverage + rewrite pyav fallback for torchvision >0.26 (#3588) Steven Palma 2026-05-12 16:59:11 +02:00
  • 01e2228b24 feat(smolvla2): per-component prompt dropout + augmented training script Pepijn 2026-05-12 15:52:32 +02:00
  • e963e5a0c4 RL stack refactoring (#3075) Khalil Meftah 2026-05-12 15:49:54 +02:00
  • c36de3a3e8 fix(smolvla2): enqueue full chunk via predict_action_chunk Pepijn 2026-05-12 15:27:23 +02:00
  • cbfaf2c544 feat(smolvla2): action-dispatch counter + tighter gibberish filter Pepijn 2026-05-12 15:22:36 +02:00
  • d0278ea093 feat(smolvla2): render state panel in autonomous mode too Pepijn 2026-05-12 15:16:28 +02:00
  • 15f6b08b0e fix(smolvla2): use canonical _strip_lerobot_blocks for inference msgs Pepijn 2026-05-12 15:07:39 +02:00
  • fc715db4a3 fix(smolvla2): coerce str content to list-of-blocks for chat template Pepijn 2026-05-12 15:01:53 +02:00
  • fe4bd2b6ba fix(smolvla2): pass flat batch dict to preprocessor (no manual wrap) Pepijn 2026-05-12 14:54:48 +02:00
  • 3f7436ff8a fix(smolvla2): use TransitionKey enum (not .value) as transition keys Pepijn 2026-05-12 14:50:22 +02:00
  • 992d13d4e9 fix(smolvla2): use build_inference_frame for raw robot observations Pepijn 2026-05-12 14:47:59 +02:00
  • afe40a016b fix(smolvla2): wrap robot obs in EnvTransition before preprocessor Pepijn 2026-05-12 14:44:24 +02:00
  • 41095e3cc3 fix(smolvla2): instantiate CameraConfig subclasses from JSON dicts Pepijn 2026-05-12 14:39:28 +02:00
  • e0fa957569 fix(smolvla2): eagerly import robot submodules before get_choice_class Pepijn 2026-05-12 14:31:58 +02:00
  • c661d81409 fix(smolvla2): use RobotConfig.max_relative_target, drop --max_action_norm Pepijn 2026-05-12 11:41:57 +02:00
  • 26ff40ddd7 chore(deps): cap torch ceiling at <2.12, pin Linux wheels to cu128 (#3570) Steven Palma 2026-05-11 19:47:55 +02:00
  • 6d269b28c8 docs(omx): adding some examples and scripts (#3566) Maxime Ellerbach 2026-05-11 15:36:32 +02:00
  • b607c8458e docs: add policy & compute guide (#3534) Steven Palma 2026-05-11 15:19:12 +02:00
  • 9e83510c99 fix(datasets): close file handle on VideoDecoder init failure in cache (#3542) Jash Shah 2026-05-10 08:30:37 -07:00
  • 1c9fbba9a9 chore(evo1): align with policy contribution guide conventions pr/3545 javadcc_mac 2026-05-10 23:14:23 +08:00
  • 6a1b5ceb9d Merge remote-tracking branch 'upstream/main' into codex/add-evo1-policy javadcc_mac 2026-05-10 22:48:17 +08:00
  • daa4c4dd30 chore(lock): regenerate uv.lock for evo1 extra javadcc_mac 2026-05-10 22:43:26 +08:00