v0.6.0

What's Changed

• fix(datasets): set batch_size in v2.0->v2.1 converter fake config by @shuheng-liu in #313
• feat(envs): make LIBERO sim control frequency configurable by @shuheng-liu in #312
• fix(pi07_paligemma): align low-level policy with pi07 low-level by @shuheng-liu in #315
• feat(train): log LIBERO grid-summary eval videos to wandb by @shuheng-liu in #316
• fix(pi07_paligemma): decouple n_action_steps from chunk_size by @shuheng-liu in #317
• ci(gpu): point LIBERO at bundled package assets in nightly test by @shuheng-liu in #322
• fix(policies): decouple n_action_steps from chunk_size by @shuheng-liu in #320
• feat(datasets): mixed-frequency training + fps metadata for pi07 by @shuheng-liu in #323
• fix(scripts): support action_freq=None in fit_fast_tokenizer manual sampler by @shuheng-liu in #324
• fix(datasets): allow action_freq=None in WeightedDatasetMixture signature by @shuheng-liu in #325
• fix(datasets): resolve per-episode task_index from parquet for speed bucketing by @shuheng-liu in #326
• fix(scripts): pass format="mp4" to wandb.Video for grid_summary logging by @shuheng-liu in #327
• fix(envs): zero-fill missing camera slots at eval; libero num_cams by @shuheng-liu in #328
• fix(logging): collapse per-dataloader sampler-save logs at checkpoint by @shuheng-liu in #329
• fix(datasets): dedup v2.0 dataset format warning across mixture loads by @shuheng-liu in #330
• chore: filter spurious pytorch 2.10 pin_memory deprecation warning by @shuheng-liu in #331
• fix(datasets): switch default video decoder to torchcodec (bump 0.4->0.10) by @shuheng-liu in #332
• feat(envs): inject subgoal images into LIBERO eval by @shuheng-liu in #333
• ci: skip redundant PR runs when push covered the same tree by @shuheng-liu in #334
• feat(pi07_paligemma): skip saved normalization buffers on load by @shuheng-liu in #337
• feat(policies): per-dataset normalization + skip-stats safetensors flag by @shuheng-liu in #336
• fix(configs+tests): unblock dev_config smoke and merge-collision test failures by @shuheng-liu in #339
• feat(policies): elevate skip_normalization_weights to all policies by @shuheng-liu in #340
• fix(scripts): inject dataset_index in inference smoke obs by @shuheng-liu in #342
• chore(ci): make extract-claude-lessons fail gracefully by @shuheng-liu in #343
• fix(policies): mask zero-pad action dims in flow-matching MSE by @shuheng-liu in #344
• feat(policies): per-(robot_type, control_mode) normalization heads by @shuheng-liu in #347
• feat(scripts): per-(robot_type, control_mode) normalization in fit_fast_tokenizer by @shuheng-liu in #348
• perf(pi07_paligemma): skip subgoal image tower when all-padded by @shuheng-liu in #351
• feat(scripts): inference-speed benchmark + fix pi07_paligemma registration by @shuheng-liu in #352
• fix(scripts): commit wandb step after all per-step logging by @shuheng-liu in #354
• fix(scripts): compact LIBERO eval logging, drop verbose video_paths by @shuheng-liu in #355
• build(deps): numpy 2.x / gymnasium / CMake 4 LIBERO, drop libero+urdf conflict by @shuheng-liu in #356
• fix(datasets): dual-split feature-mapping collision + norm-distribution diagnostic by @shuheng-liu in #359
• feat: dataset provenance fields + pi07_paligemma outlier warning by @shuheng-liu in #357
• fix(deps): bump pyopengl-accelerate to 3.1.10 for numpy 2.x (libero) by @shuheng-liu in #361
• feat(diagnose_norm): per-dim histograms vs N(0,1), replace heatmaps by @shuheng-liu in #362
• fix(datasets): pool norm stats over selected episodes + add excluded_episodes by @shuheng-liu in #363
• test: cover skip_normalization_weights with N!=M per-norm-key heads by @shuheng-liu in #364
• fix: history_state_drop masks history, keeps current state, zeros post-norm by @shuheng-liu in #365
• fix: guard zero-variance dims in Normalize + pad-aware outlier warning by @shuheng-liu in #366
• Fix fake_tensor_training.py crash on torch 2.10 (FakeTensor .to swap) by @shuheng-liu in #367
• Outlier warning always reaches wandb (gather to rank 0); port to pi07 by @shuheng-liu in #368
• feat(pi07_paligemma): custom CUDA block-causal flash attention by @WilliamYue37 in #358
• feat(pi07_paligemma): per-(robot_type, control_mode) state/action projections by @shuheng-liu in #370
• feat(pi07): per-(robot_type, control_mode) state/action projections by @shuheng-liu in #371
• test(pi07): use one import style in per-group projection test by @shuheng-liu in #372
• feat(train): aggregate val loss per (dataset, control_mode), parallelize by @shuheng-liu in #374
• chore(claude): learn from #374 by @claude[bot] in #375
• feat(datasets): allow per-dataset override of mixture val_split_ratio by @shuheng-liu in #377
• build(libero): robosuite 1.5.2 + mink override for RoboCasa coexistence by @shuheng-liu in #378
• feat(envs): add RoboCasa365 simulated-eval environment alongside LIBERO by @shuheng-liu in #386
• feat(envs): auto-download RoboCasa assets to a relocatable external store by @shuheng-liu in #387
• chore(claude): learn from #387 by @claude[bot] in #388
• feat(datasets): add read-only support for LeRobot v3.0 datasets by @shuheng-liu in #391
• docs: correct stale optional-dependency references and RoboCasa eval docs by @shuheng-liu in #392
• Fix logging of success rate on robocasa eval by @akshay18iitg in #390
• perf(train): free the CUDA allocator cache before in-training eval by @shuheng-liu in #393
• fix(ci): install robocasa extra in nightly GPU job for RoboCasa test by @shuheng-liu in #395
• feat(datasets): default to main/master when no codebase-version tag by @shuheng-liu in #397
• feat(envs): render all cameras in eval rollout videos by @shuheng-liu in #398
• fix(train): broadcast output_dir so all ranks share one checkpoint tree by @shuheng-liu in #396
• perf(eval): stream grid summary video to bound memory by @shuheng-liu in #399
• fix(ci): use HF_TOKEN env var instead of flaky hf auth login step by @shuheng-liu in #400
• fix(eval): honor max_episodes_rendered and grid_size from EvalConfig by @shuheng-liu in #401
• fix(envs): pin MuJoCo EGL render to per-rank GPU for multi-GPU sim eval by @shuheng-liu in #402
• feat(config): add wandb option to skip eval video logging by @shuheng-liu in #405
• fix(envs): register nvidia EGL ICD so RoboCasa sim eval renders on GPU by @shuheng-liu in #404
• fix(policies): single-head norm fallback for older checkpoints by @shuheng-liu in #406
• chore: bump version to 0.6.0 by @WilliamYue37 in #407

Full Changelog: v0.5.0...v0.6.0