All pull requests

Pull requests

• fix(eval): collect predicted latents through PEFT wrapped model

  #5092·HaomingSong opened Oct 25, 2026Collaborator·

  ·
  1

• fix(scripts): warn when fine-tuning replaces pretrained processor stats with dataset stats

  tests
  #5091·akushonkamen opened Oct 25, 2026·2 of 5 tasks·

  ·
  1

• docs(openarm): nest recording flags under --dataset

  documentation
  #5090·ZainnQureshii opened Oct 25, 2026·2 of 5 tasks·

  ·
  1

• docs(eo1): update chunk_size and n_action_steps defaults to 16

  documentation
  #5089·ZainnQureshii opened Oct 25, 2026·2 of 5 tasks·

  ·
  1

• Fix SO-ARM101 hardware safety gate integration

  robotstests
  #5088·vantek988 opened Oct 25, 2026·

  ·
  1

• feat(train): support PEFT under sharded training (FSDP2, HSDP)

  configurationdocumentationpoliciestests
  #5086·HaomingSong opened Oct 25, 2026Collaborator·2/2·

  ·
  1

• fix(processor): preserve squeeze_batch_dim across save/load

  processortests
  #5085·not-heavychevy opened Oct 24, 2026·3 of 5 tasks·

  ·

• fix(xvla): apply the LIBERO env processors to LIBERO only and check the action conversion

  evaluationpolicies
  #5084·HaomingSong opened Oct 24, 2026Collaborator·

  ·
  1

• Add SO-ARM101 hardware safety gate

  robotstests
  #5083·vantek988 opened Oct 24, 2026·

  ·

• feat(rl): add QF3 filtered-gradient updates for a state-only flow actor

  #5081·Hadrien-Cornier opened Oct 24, 2026Contributor·5 tasks·

  ·
  1

• feat(datasets): add Image layouts and read APIs to Lance

  #5079·yanghua opened Oct 24, 2026Contributor·1 of 5 tasks·

  ·
  1

• fix(datasets): escape apostrophes in ffconcat paths

  datasettests
  #5078·mangeshraut712 opened Oct 23, 2026·3 of 5 tasks·

  ·
  1

  1

• fix(factory): keep the output features of a checkpoint at evaluation

  policies
  #5077·HaomingSong opened Oct 23, 2026Collaborator·1/2·

  ·
  3

• docs(so101): clarify servo orientation before assembly

  documentation
  #5076·kimjune01 opened Oct 23, 2026·2 of 5 tasks·

  ·

• docs(hilserl): correct leader-arm, env type, reward label and crop guidance

  documentation
  #5074·Shavvimal opened Oct 23, 2026·2 of 5 tasks·

  ·

• fix(rl): clamp out-of-range discrete actions in the SAC discrete critic

  tests
  #5073·Shavvimal opened Oct 23, 2026·3 of 5 tasks·

  ·

• fix(processor): move reward classifier inputs to the classifier device

  processortests
  #5072·Shavvimal opened Oct 23, 2026·2 of 5 tasks·

  ·

• chore(ci): cut Fast Tests time by removing redundant work and redundant cases in the slowest tests

  datasettests
  #5070·Maximellerbach opened Oct 23, 2026Member·2/2·2 of 5 tasks·

  ·
  4

• chore(ci): Make CI great again

  CIdocumentationgithub_actionstests
  #5069·Maximellerbach opened Oct 23, 2026Member·1/2·

  ·
  1

• feat(train): opt-in DataLoader worker recycling to bound host memory

  configurationtests
  #5068·kkipngenokoech opened Oct 23, 2026Contributor·3 of 5 tasks·

  ·
  3

• perf(smolvla): precompute cross-attention expert K/V once per chunk

  policiestests
  #5067·usersforsomebody opened Oct 23, 2026·4 of 5 tasks·

  ·
  1

• fix(datasets): reject mislabeled episodes in v2.1 to v3.0 conversion

  tests
  #5066·sypsyp97 opened Oct 23, 2026·3 of 5 tasks·

  ·
  1

• fix(training): enable CPU collectives for CUDA FSDP offload

  #5065·manfromnowhere143 opened Oct 23, 2026·1 of 5 tasks·

  ·

• fix(training): preserve per-rank RNG in distributed checkpoints

  tests
  #5064·manfromnowhere143 opened Oct 22, 2026·4 of 5 tasks·

  ·

• fix(motors): flip homing offset sign on inverted Dynamixel motors

  robotstests
  #5063·akushonkamen opened Oct 22, 2026·2 of 5 tasks·

  ·
  1