Issues

• Near-zero standard deviation on a constant joint silently blows up normalized values at rollout

  bugdatasetdocumentationpoliciestraining
  huggingface/lerobot#5289 · Filter by author SashaBajzek opened on Nov 8, 2026

• ReplayBuffer(optimize_memory=True) returns the wrong next_state at episode and intervention boundaries

  datasettraining
  huggingface/lerobot#5284 · Filter by author SofianElmotiem opened on Nov 7, 2026
  

• Modernize Type Param syntax to PEP 695

  enhancementexamplesprocessor
  huggingface/lerobot#5242 · Filter by author FrancescoPezzella opened on Nov 4, 2026

• OpenArm connect rewrites Damiao zeros; left gripper limit is not mirrored

  bugexamplesrobotssensorsteleoperators
  huggingface/lerobot#5241 · Filter by author ogawakonatsu-hue opened on Nov 4, 2026
  2

• Pending private security advisory awaiting maintainer review

  huggingface/lerobot#5233 · Filter by author r3vpwnx opened on Nov 3, 2026

• [Windows] Unstable Concurrent Capture from Two USB 2.0 UVC Cameras (One Camera Drops to ~1 FPS)

  bugconfigurationperformancepoliciessensorsteleoperatorstests
  huggingface/lerobot#5193 · Filter by author zhu57320041-alt opened on Nov 1, 2026

• lerobot-train profile on one H200 (ACT, SmolVLA, Diffusion): where step time and memory go

  configurationdatasetenhancementevaluationexamplesperformancepoliciesprocessorquestionsensorstraining
  huggingface/lerobot#5189 · Filter by author arun-elr opened on Nov 1, 2026

• [DATA-04] Frame addressing: use integer frame indices for video (measurements from 38 Hub datasets and barcoded synthetic sets)

  bugdatasetdocumentationexamplesperformancepoliciessensorstraining
  huggingface/lerobot#5187 · Filter by author arun-elr opened on Nov 1, 2026

• LiberoEnv defaults to 360 px observations but the SmolVLA LIBERO checkpoints were trained at 256 px (0/20 vs 6/20 on two LIBERO-10 tasks)

  configurationevaluationpoliciesprocessorsensorstestsvisualization
  huggingface/lerobot#5185 · Filter by author jdubray opened on Oct 31, 2026
  8

• FineART: observation.state[t] equals action[t-1] exactly in every frame

  bugpolicies
  huggingface/lerobot#5184 · Filter by author ebeirne opened on Oct 31, 2026
  2

• make_policy leaves the resume checkpoint in cfg.pretrained_path when construction fails

  bugconfigurationexamplespoliciesteststrainingvisualization
  huggingface/lerobot#5176 · Filter by author JonnyFaeth opened on Oct 31, 2026
  3

• 2x faster flow-policy inference with a one-pass Gaussian readout

  bugenhancementevaluationexamplesperformancepoliciesprocessorsimulationteleoperatorstests
  huggingface/lerobot#5173 · Filter by author Hadrien-Cornier opened on Oct 31, 2026

• Implement IMPRINT (Memory Without Curation): latent episode memory for policies

  datasetdocumentationenhancementpoliciessensorsteststraining
  huggingface/lerobot#5172 · Filter by author Hadrien-Cornier opened on Oct 30, 2026

• Streaming-WAM: 2.9x faster warm inference for Fast-WAM, no pause at chunk refresh

  enhancementperformancepoliciestests
  huggingface/lerobot#5164 · Filter by author Hadrien-Cornier opened on Oct 30, 2026

• [Performance] pi0.5: the denoise loop is 86% of a bf16 chunk; static prefix KV + CUDA-graph denoise?

  bugenhancementperformancepoliciessensors
  huggingface/lerobot#5148 · Filter by author arun-elr opened on Oct 29, 2026
  5

• remove_feature leaves stale per-episode stats/video columns, causing delete_episodes/split_dataset to fail in aggregate_stats

  bugdatasetsensors
  huggingface/lerobot#5138 · Filter by author Teddy-Liao opened on Oct 29, 2026

• Example proposal: converting rosbag2/MCAP recordings to LeRobotDataset v3.0 (topic-to-feature mapping recipe)

  datasetdocumentationenhancementexamplesperformancequestionteleoperatorstests
  huggingface/lerobot#5130 · Filter by author kittimzhe opened on Oct 28, 2026
  4

• PI0 / PI05 / SmolVLA put remaining cameras in the wrong slots when one camera is missing

  bugconfigurationevaluationexamplespoliciessensors
  huggingface/lerobot#5118 · Filter by author chaoshengsc opened on Oct 27, 2026
  5

• pi0: padded end-of-episode actions are included in the training loss

  bugconfigurationdatasetexamplespoliciesteststraining
  huggingface/lerobot#5105 · Filter by author johyeongseob opened on Oct 26, 2026

• [Discussion] Retrofitting legacy industrial arms with VLA — what's the practical path?

  performancepoliciesprocessorsimulationteleoperatorstests
  huggingface/lerobot#5102 · Filter by author bkbk-robot opened on Oct 26, 2026
  6

• [Discussion] Has the bf16 update-erasure paper been evaluated or reproduced in LeRobot?

  evaluationtraining
  huggingface/lerobot#5099 · Filter by author Muyun99 opened on Oct 26, 2026
  2

• SmolVLA fine-tuning on ~100 episodes: policy output collapses toward near-constant actions (gripper span 1.4° vs 32° in demo data)

  bugconfigurationdatasetdependenciesevaluationexamplespoliciesprocessorsensors
  huggingface/lerobot#5082 · Filter by author bkbk-robot opened on Oct 24, 2026
  5

• Add QF3 filtered-gradient updates for a state-only flow actor

  bugenhancementevaluationexamplesperformancepoliciessensorssimulationtrainingvisualization
  huggingface/lerobot#5080 · Filter by author Hadrien-Cornier opened on Oct 24, 2026

• concatenate_video_files fails for input paths containing apostrophes (tested fix included)

  bugdatasetsensorstests
  huggingface/lerobot#5075 · Filter by author JonnyFaeth opened on Oct 23, 2026

• Koch leader calibration fails with "Negative values are not allowed": homing offset has the wrong sign on motors in reverse drive mode

  bugpoliciesrobotsteleoperatorstests
  huggingface/lerobot#5060 · Filter by author diegoakel opened on Oct 22, 2026
  4