admin/lerobot
1
0
Fork 0
mirror of https://github.com/huggingface/lerobot.git synced 2026-05-23 20:50:02 +00:00
Code Issues Packages Projects Releases Wiki Activity

nit

Browse Source
This commit is contained in:
Pepijn
2026-04-01 18:04:03 +02:00
parent 58bd11caf3
commit d028978552
2 changed files with 1 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/source/umi_pi0_relative_ee.mdx
-25
View File
@@ -138,31 +138,6 @@ Inference:
IK → joint targets → robot IK → joint targets → robot
``` ```
## Manual mode (without derive_state_from_action)
If your dataset already has `observation.state` (or you want to add it separately), you can skip `derive_state_from_action` and use relative actions + relative state independently:
```bash
# Recompute stats
lerobot-edit-dataset \
--repo_id <your_dataset> \
--operation.type recompute_stats \
--operation.relative_action true \
--operation.relative_state true \
--operation.state_obs_steps 2 \
--operation.chunk_size 50 \
--operation.relative_exclude_joints "['gripper']"
# Train
lerobot-train \
--dataset.repo_id=<your_dataset> \
--policy.type=pi0 \
--policy.use_relative_actions=true \
--policy.use_relative_state=true \
--policy.state_obs_steps=2 \
--policy.relative_exclude_joints='["gripper"]'
```
## References ## References
- [UMI: Universal Manipulation Interface](https://umi-gripper.github.io) — Chi et al., 2024. Defines relative trajectory actions. - [UMI: Universal Manipulation Interface](https://umi-gripper.github.io) — Chi et al., 2024. Defines relative trajectory actions.
src/lerobot/policies/pi0_fast/processor_pi0_fast.py
+1 -1
View File
@@ -136,7 +136,7 @@ def make_pi0_fast_pre_post_processors(
# Pi0Fast order: relative → normalize → tokenize → model → unnormalize → absolute # Pi0Fast order: relative → normalize → tokenize → model → unnormalize → absolute
# This matches pi0/pi0.5: RelativeActionsProcessorStep runs first on raw absolute actions, # This matches pi0/pi0.5: RelativeActionsProcessorStep runs first on raw absolute actions,
# caching the raw state. NormalizerProcessorStep then normalizes the raw relative actions, # caching the raw state. NormalizerProcessorStep then normalizes the raw relative actions,
# so the normalizer (and action tokenizer) sees delta values — relative stats are required. # so the normalizer (and action tokenizer) sees delta values, relative stats are required.
# NOTE: RelativeActionsProcessorStep only modifies the action in the transition; it reads # NOTE: RelativeActionsProcessorStep only modifies the action in the transition; it reads
# state from the observation but does not change it. NormalizerProcessorStep still runs # state from the observation but does not change it. NormalizerProcessorStep still runs
# before Pi0FastPrepareStateAndLanguageTokenizerProcessorStep, so the state tokenizer # before Pi0FastPrepareStateAndLanguageTokenizerProcessorStep, so the state tokenizer