v0.5.0

What's Changed

• feat: support observation history in standard data format by @shuheng-liu in #154
• Adding AWR algorithm for pi0 by @akshay18iitg in #110
• Adding high level memory by @akshay18iitg in #157
• Feat/pi05 mem by @WilliamYue37 in #153
• feat: support deferred video attachment in LeRobotDataset by @shuheng-liu in #158
• fix: placeholder stats for deferred video keys respect channel axis by @shuheng-liu in #164
• feat: support spec. start time of attached video by @shuheng-liu in #160
• perf: parallelize ffmpeg trims in segment_lerobot_dataset by @bananaSnail in #166
• Adding script to parse subtask.json and add response field to parquet… by @akshay18iitg in #163
• feat: attach metadata and expose optional standard data format keys by @shuheng-liu in #168
• perf(pi05): 2.9x LIBERO training throughput via DDP + fused AdamW by @shuheng-liu in #176
• feat: per-dataset validation loss reporting by @shuheng-liu in #169
• Maintaining consistency for pi05 with paligemma pretraining by @akshay18iitg in #161
• fix: mixture-weighted validation aggregate + per-name dup suffix by @shuheng-liu in #189
• Feat/pi07 paligemma by @akshay18iitg in #167
• feat: overlay-repo support for LeRobot v2.x datasets by @shuheng-liu in #183
• feat(policies): add authentic pi06 policy with Gemma 3 4B backbone by @shuheng-liu in #178
• chore(ci): add CLAUDE.md and three Claude integration workflows by @shuheng-liu in #199
• fix(ci): grant Claude workflows bypassPermissions for tool calls by @shuheng-liu in #207
• feat: replace V-JEPA2 with space-time SigLIP video encoder in pi05_mem by @shuheng-liu in #171
• feat: surface robot_type and control_mode in standard data format by @shuheng-liu in #204
• feat(ci): pin Claude workflows to Opus 4.7 1M, allow test-skip on doc fixes by @shuheng-liu in #208
• ci: skip pi07_paligemma low-level test (broken at import) to unblock CPU CI by @shuheng-liu in #209
• docs(claude): mention Claude integration workflows in CLAUDE.md by @shuheng-liu in #206
• fix(pi05): DDP fp32 Adam state + SDPA / grad-ckpt opt-ins (#181, #177) by @shuheng-liu in #182
• fix(ci): gate extract-claude-lessons on head branch, not PR author by @shuheng-liu in #212
• refactor(policies): remove init_strategy by @shuheng-liu in #188
• fix(build): pin cmake<4 in egl-probe PEP 517 build isolation by @shuheng-liu in #216
• ci: ignore broken pi07 test in GPU CI; dedupe TODO in cpu_test by @shuheng-liu in #217
• feat(scripts): annotate_subtasks.py — VLM subtask labelling for dataset mixtures by @WilliamYue37 in #215
• feat(datasets): override + require non-empty robot_type/control_mode by @shuheng-liu in #220
• feat(configs): support $ref include in JSON configs by @shuheng-liu in #222
• fix(pi06): align with π0.6 model card paper by @shuheng-liu in #235
• feat(grounding): add codec + ensure_loc_tokens for pi05/pi06 by @shuheng-liu in #237
• test(pi07): drop VJEPA2 test + fix cpu_test.yml self-race (#234) by @shuheng-liu in #232
• Deprecate DatasetConfig.val_split_ratio in favor of mixture-level by @shuheng-liu in #225
• feat(pi0, pi06): add SDPA attention + gradient checkpointing options by @shuheng-liu in #214
• ci: gate workflows on org membership + allow claude[bot] by @shuheng-liu in #246
• test(pi06): verify pi06 ckpt loads + matches gemma-3-4b-pt by @shuheng-liu in #245
• ci: remove concurrency group from cpu_test workflow by @shuheng-liu in #248
• feat(pi07): π0.7 policy with Gemma 3 backbone and SpaceTime SigLIP video encoder by @shuheng-liu in #244
• ci: review [claude-fix] commits; bar bots from triggering @claude fix by @shuheng-liu in #251
• refactor(pi07): rename low_level_planner -> low_level (it's not a planner) by @shuheng-liu in #250
• chore(claude): learn from #242 by @claude[bot] in #243
• fix(pi07): mask padded history; preserve current state token by @shuheng-liu in #253
• feat(annotate_subtasks): add Gemini Robotics-ER 1.6 as a model option by @WilliamYue37 in #255
• docs(annotate_subtasks): document Gemini Robotics-ER 1.6 support by @WilliamYue37 in #256
• feat(annotate_subtasks): 448x448 frames; select camera0 from standard mapping by @WilliamYue37 in #257
• feat(datasets): add tolerance_s and skip_timestamp_check config knobs by @shuheng-liu in #258
• feat(annotate_subtasks): only downsample frames larger than target_size by @WilliamYue37 in #261
• Fix/pi07 paligemma fixes by @akshay18iitg in #260
• test(pi07): verify pi07-{high,low}-untrained + drop expert in high by @shuheng-liu in #263
• fix(pi05_mem): mask padded history; preserve current state token; emit at inference by @shuheng-liu in #254
• refactor(video_encoder): unify SigLIP video encoder; allow variable T by @WilliamYue37 in #268
• refactor(pi07_paligemma): layout-agnostic ContextItem prefix/suffix by @WilliamYue37 in #262
• Fixing SpaceTime attention block in siglip by @akshay18iitg in #271
• feat(pi07,fsdp): enable FSDP-FULL_SHARD for full unfreeze + profile_step audit + ZeRO-2 vs FSDP matrix by @shuheng-liu in #273
• refactor(configs): consolidate n_obs_history/history_interval; pi07 num_steps→5 by @shuheng-liu in #274
• chore(claude): learn from #273 by @claude[bot] in #275
• test: remove tests requiring private TensorAuto/*-untrained repos (#272) by @WilliamYue37 in #279
• test(pi07): set vocab size + FSDP MixedPrecision in FSDP-wrap test (#281) by @shuheng-liu in #282
• feat(scripts): add annotate_mistakes.py for per-frame mistake labels by @WilliamYue37 in #280
• chore(datasets): collapse skip-timestamp-check warning to once-per-process by @shuheng-liu in #278
• test(pi07): unblock FSDP-wrap GPU test on NCCL 2.27.5 single rank (#283) by @shuheng-liu in #285
• ci(gpu) + test(pi07): drop temp push trigger + scope CUDA drain (cleanup of #285) by @shuheng-liu in #286
• fix(train): re-seed new ranks deterministically when resuming on more GPUs by @shuheng-liu in #180
• fix(datasets): NaN-tolerant aggregate_feature_stats + load-time WARN by @shuheng-liu in #288
• ci: make Claude PR Review non-blocking when API limit hits by @shuheng-liu in #291
• ci: move Claude PR Review continue-on-error to step level by @shuheng-liu in #293
• feat(policies): support training and selecting FAST tokenizer by @shuheng-liu in #290
• log(metrics): show exact step/samples + bump loss precision to 6dp by @shuheng-liu in #292
• feat(train): log per-rank HBM around LIBERO eval boundary by @shuheng-liu in #294
• fix(metrics): only report L1/Accuracy when policy returns them by @shuheng-liu in #297
• feat(datasets): bucket speed by duration-seconds, not native frames by @shuheng-liu in #295
• feat(eval): configurable pi07 metadata on EnvConfig by @shuheng-liu in #296
• fix(configs): remove invalid fields from pi07 libero smoke config by @shuheng-liu in https://github.com/TensorAuto/Open...

v0.4.0

What's Changed

• Added training and accelerate configs for reproducing libero results with PI05 by @WilliamYue37 in #118
• Adding warning when discrete action is truncated and increased dicret… by @akshay18iitg in #119
• Deprecating loss type mapping by @akshay18iitg in #120
• Feat/value response by @akshay18iitg in #109
• Feat: added a script to convert human demos (ego or exo) to LeRobot datasets by @WilliamYue37 in #121
• fix: guard type of states names in info.json by @shuheng-liu in #123
• Rename grounding to vqa by @WilliamYue37 in #124
• Add support for collecting human data with the RecordHuman PICO VR app by @WilliamYue37 in #126
• feat: add script to segment lerobot dataset by @shuheng-liu in #127
• Keep asset video by default for rerun visualization by @shuheng-liu in #135
• fix: upsert dataset_feature_name_mapping if specified with dataset by @shuheng-liu in #139
• Adding value function visualization by @akshay18iitg in #137
• fix get item by @akshay18iitg in #144
• Updating RL results by @akshay18iitg in #146
• Feat/pi05 continuous state by @WilliamYue37 in #145
• Fixes/libero rollout by @akshay18iitg in #141
• Support empty mixture weights (default to lengths) by @shuheng-liu in #149
• Support convert checkpoint from any directory by @shuheng-liu in #148
• feat: add tqdm bar to the dataset segment script by @shuheng-liu in #150
• Adding robocasa server and client for batch inference by @akshay18iitg in #151
• Remove temporal grouping from dataset pipeline by @shuheng-liu in #152
• Chore/bump v0.4.0 by @WilliamYue37 in #155

Full Changelog: 0.3.0...v0.4.0

v0.3.0

What's Changed

• Adding converstion script from ros format to lerobot format by @akshay18iitg in #70
• Only calculating response loss when response_predict config is set t… by @akshay18iitg in #105
• Changed inference.py to compile policy.sample actions by @WilliamYue37 in #106
• feat: Added training-time real time chunking (RTC) by @WilliamYue37 in #101
• Switch torch compile to use max-autotune mode by @WilliamYue37 in #111
• feat: support setting joint names for each state dim in URDF viz by @shuheng-liu in #108
• Feat: support ONNX export/inference + TensorRT inference by @shuheng-liu in #103
• fix(model): Compile policy.model.sample_actions instead of policy.sample_actions by @WilliamYue37 in #114
• Adding server grpc with async code by @akshay18iitg in #112
• Chore: Bump versioning to 0.3.0 by @WilliamYue37 in #115

Full Changelog: v0.2.0...0.3.0

Release v0.2.0

What's Changed

• URDF support in dataset visualization by @shuheng-liu in #67
• Fixing calculation of prefix_offsets in denoise step of pi05 policy by @akshay18iitg in #72
• Feat: Added a PR checklist that gets enforced by CI by @WilliamYue37 in #75
• Added optional validation dataset split during training by @WilliamYue37 in #77
• Fix gymnasium import by @shuheng-liu in #79
• Added Github Actions for GPU pytests that run nightly by @WilliamYue37 in #62
• Adding Sub task planner for training and inferencing in pi05 by @akshay18iitg in #65
• Move GPU pytests from AWS us-west-1 to us-west-2 by @WilliamYue37 in #80
• Updating dependencies to use torch 2.8 by @akshay18iitg in #83
• Fix Docker build by update uv sync command by @WilliamYue37 in #86
• Add Regression Tests by @WilliamYue37 in #85
• Remove response from dummy observations for pi0.5 inference by @shuheng-liu in #91
• Added Badges to README by @WilliamYue37 in #92
• Freezing discrete lm head and embedding matrix by @akshay18iitg in #95
• Add cache to regression test by @akshay18iitg in #96
• gRPC server and client by @WilliamYue37 in #90
• Fixes flaky regression tests caused by race conditions in metadata download by @WilliamYue37 in #98
• fix: properly implementPatchedGemmaRMSNorm:extra_repr by @shuheng-liu in #100
• chore(dependencies): bump opentau to v0.2.0 by @WilliamYue37 in #99

Full Changelog: v0.1.2...v0.2.0

v0.1.2 - Improved dependencies

What's Changed

• Supports visualization of LeRobot datasets using rerun and removes the flask-based visualizer by @shuheng-liu in #58
• Switch from opentau-transformers to transformers by @WilliamYue37 in #61
• updated pyproject version by @WilliamYue37 in #63

Full Changelog: v0.1.1...v0.1.2

v0.1.1 - Bug Fixes

What's Changed

• Add a PyPI publisher workflow to github actions by @WilliamYue37 in #49
• Fixes gradient stepping in training script by @shuheng-liu in #53
• Fixes CPU pytests by @shuheng-liu in #54
• Fix issue with accelerate looking up the metadata of the transformers package by @WilliamYue37 in #56
• Added new entry point commands for eval and export to onnx by @WilliamYue37 in #57
• update pypi version by @WilliamYue37 in #59

Full Changelog: v0.1.0...v0.1.1

Initial Public Release

v0.1.0

Initial public release