[release/0.2.1] Change the prompt for the finetuning

[pulkitg01]

Summary

Align static apple prompt

Detailed description

• Reason: align the simulated static apple workflow with the same prompt used for the real setup.
• Changed the GR00T LeRobot training config prompt to move the apple to the plate.
• Updated the closed-loop evaluation config to use the same prompt.
• Updated the static apple policy-training docs snippet so users finetune with the matching prompt.
• Impact: training and evaluation now use consistent language conditioning across sim and real workflows.

[isaaclab-review-bot]

Review Summary

This PR simplifies the language instruction prompt from a detailed description to a more concise task statement across 3 configuration files and documentation.

Changes Reviewed

File	Change
docs/pages/example_workflows/static_apple/step_3_policy_training.rst	Updated documentation prompt
isaaclab_arena_gr00t/lerobot/config/g1_static_apple_config.yaml	Updated training config prompt
isaaclab_arena_gr00t/policy/config/g1_static_apple_gr00t_closedloop_config.yaml	Updated inference config prompt

Observations

✅ Consistency: The prompt change is applied consistently across all relevant files (documentation, LeRobot training config, and closed-loop policy config).

✅ Commit hygiene: Good to see the follow-up commit removing the trailing period for cleaner prompt formatting.

Clarification Request

The prompt is changing from:

"Pick up the apple from the shelf and place it onto the plate on the same shelf next to it."

to:

"move the apple to the plate"

Question: Could you confirm this matches the prompt used during the finetuning of checkpoint-20000? Language-conditioned policies can be sensitive to prompt wording, so ensuring the inference prompt matches the training prompt is important for optimal performance.

---

This is an automated review. Please address any questions before merging.

---

Update (4ab975d): Reviewed incremental changes from 23c7a78 to 4ab975d.

Additional Changes in This Push

This update contains significant documentation improvements across the static apple workflow and related teleoperation guides:

Documentation restructuring (step_1_environment_setup.rst):

• Condensed the verbose code walkthrough into a cleaner component table format
• Moved implementation details to a concise "Environment Composition" section
• Removed inline code snippets in favor of pointing to the source file

CloudXR network requirements (multiple files):

• Added consistent .. important:: admonition at the top of teleoperation steps to highlight network requirements upfront
• Removed scattered inline network setup notes that were redundant with the new consolidated requirement

Training workflow updates (step_3_policy_training.rst):

• Changed recommended GPU config from 8x L40s to single RTX 6000 Ada with adjusted batch size (12 vs 96)
• Added dropdown for pre-generated dataset download from HuggingFace
• Added explicit host path guidance for DATASET_DIR and MODELS_DIR outside Docker

Evaluation workflow updates (step_4_evaluation.rst):

• Added dropdown for pre-trained model download from HuggingFace
• Added new "Parallel Environments Evaluation" section for multi-env testing
• Added note about Kit permission errors and stale cache issues
• Improved troubleshooting guidance for action shape mismatches

Minor fixes:

• Updated GR00T repo links to point to specific commit 4b1dca9d88d2a0b9ea5a65aa61c82ff89f5c4f0e
• Fixed typo "GR00T-flavoured" → generic "Arena"
• Added headset view reset instructions for Meta Quest and PICO 4 Ultra

✅ All changes look good — documentation improvements are well-organized and the pre-recorded dataset/checkpoint downloads will make the tutorial more accessible.

---

This is an automated review. Please address any questions before merging.

[shanereetz]

LGTM 🚀

[greptile-apps]

Greptile Summary

This PR aligns the static apple workflow's language_instruction with the real-setup prompt by replacing the verbose sentence with the shorter "move the apple to the plate" across training config, evaluation config, and docs.

• Training config (g1_static_apple_config.yaml) and evaluation config (g1_static_apple_gr00t_closedloop_config.yaml) both receive the updated prompt, ensuring consistent language conditioning between sim and real.
• Documentation snippet (step_3_policy_training.rst) updates the displayed config, but line 110 retains stale prose that still characterises the old verbose instruction and will mislead readers.

Confidence Score: 4/5

Safe to merge after fixing the stale explanatory prose in the docs that still describes the old verbose instruction.

The two YAML config changes are correct and consistent. The docs file updates the displayed config snippet correctly, but line 110 still reads "describing the same-shelf placement (no walking, no second table)" — a parenthetical that was accurate for the old prompt but actively misleads readers now that the instruction is simply "move the apple to the plate".

docs/pages/example_workflows/static_apple/step_3_policy_training.rst — line 110 prose needs updating to reflect the new short prompt.

Important Files Changed

Filename	Overview
docs/pages/example_workflows/static_apple/step_3_policy_training.rst	Updates the doc snippet's language_instruction to the new short prompt, but line 110 retains stale prose that still describes the old verbose instruction.
isaaclab_arena_gr00t/lerobot/config/g1_static_apple_config.yaml	Replaces the verbose language_instruction with the shorter "move the apple to the plate" to align training config with the real setup.
isaaclab_arena_gr00t/policy/config/g1_static_apple_gr00t_closedloop_config.yaml	Updates language_instruction in the closed-loop evaluation config to match the new training prompt, ensuring consistent language conditioning.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["language_instruction\n'move the apple to the plate'"] --> B["g1_static_apple_config.yaml\n(LeRobot training config)"]
    A --> C["g1_static_apple_gr00t_closedloop_config.yaml\n(closed-loop eval config)"]
    A --> D["step_3_policy_training.rst\n(docs snippet)"]
    B --> E["GR00T N1.7 fine-tuning"]
    C --> F["Closed-loop policy evaluation"]
    D --> G["User docs\n⚠️ line 110 prose still describes old prompt"]

Loading

Comments Outside Diff (1)

1. docs/pages/example_workflows/static_apple/step_3_policy_training.rst, line 110 (link)

   Stale prose describing the old prompt

   Line 110 still reads "describing the same-shelf placement (no walking, no second table)", which was written to characterise the previous verbose instruction. Now that language_instruction is "move the apple to the plate", that parenthetical no longer reflects what the instruction says and will mislead readers comparing the static-apple config against the loco-manipulation one.

   Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

_{Reviews (2): Last reviewed commit: "Merge branch 'release/0.2.1' into pulkit..." | Re-trigger Greptile}