Skip to content

v0.14.0

Latest

Choose a tag to compare

@shuheng-liu shuheng-liu released this 14 Sep 18:25
66ecbed

What's Changed

  • docs(changelog): split [Unreleased] into 0.13.0 and 0.12.0 sections by @shuheng-liu in #519
  • feat(accel): measure the best prefix instead of assuming the paper's by @shuheng-liu in #520
  • fix(grpc): run compiled sample_actions on one dedicated thread by @cenxiaoxinxin-dotcom in #523
  • chore: drop an internal host name from a source comment by @shuheng-liu in #526
  • fix(value): learned readout token + single shared norm head by @akshay18iitg in #524
  • feat(policies): best-of-N action-chunk sampling for pi05 and pi06 by @shuheng-liu in #525
  • chore(claude): learn from #519 by @claude[bot] in #521
  • fix(tests): compare candidate arming by execution order, not file offset by @shuheng-liu in #528
  • chore(claude): point rule 2 at the GPU-test helpers from #513 by @claude[bot] in #516
  • seq_dataloader_pi0.5_ttt by @vparkar-mle in #532
  • feat(train): legacy norm-key remap, trainable-only snapshots, LIBERO task fix by @shuheng-liu in #533
  • chore(claude): learn from #533 by @claude[bot] in #534
  • chore(claude): learn from #516 by @claude[bot] in #529
  • fix(datasets): sequence-mode batches >1 collate across episode boundaries by @shuheng-liu in #535
  • fix(datasets): derive sequence_stride from action_chunk, reject mismatches by @shuheng-liu in #537
  • feat(envs): explicit RoboCasa kitchen-scene list via env.layout_and_style_ids by @shuheng-liu in #538
  • feat(pi05_ttt): inference-only diagnostics for dissecting trained TTT checkpoints by @shuheng-liu in #540
  • fix(pi05_ttt): read the alpha-scale diagnostic live off the bound config by @shuheng-liu in #541
  • feat(grpc)!: rename api-key auth to x-api-key / INFERENCE_API_KEY by @cenxiaoxinxin-dotcom in #548
  • fix(datasets): make episode sampler rank-independent by @zjn20030811 in #547
  • feat(datasets): pair demonstration+rollout episodes for one-shot imitation by @vparkar-mle in #549
  • fix(robocasa): correct the flat action layout and repair eval integrity by @shuheng-liu in #550
  • feat(xr1): port the Xiaomi-Robotics-1 policy (Qwen3-VL-4B + 36-layer DiT) by @shuheng-liu in #552
  • chore: bump version to v0.14.0 by @shuheng-liu in #554

New Contributors

Full Changelog: v0.13.0...v0.14.0