Repository navigation
feat: Phase 4.A robot_offline + action tokens (smol 256M) - #53
Merged
Merged
Conversation
Productize offline robot learning for memory-constrained robots: ActionTokenizer (OpenVLA-style bins v1), run_robot_offline over trajectories with held-out episode probes, default base SmolVLM-256M, edge rank-8 plans/throughput, convert --action-scheme bins, and catalog robot_offline_edge updates. Never actuates without a supervisor.
CI failed on I001/F401; keep isort-style blocks and drop unused trajectories_to_examples import in tests.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Phase 4.A first product gap: offline robot policy path sized for a severe memory robot that runs best on smol ~250M.
ActionTokenizer— text-tokenized actions (bins v1 default 256 over [-1,1], or continuous decimals); encode/decode; serderun_robot_offline— trajectories → tokenized examples → LoRA CE under four verbs; metrics/probes/resume/southward viarun_sft; default baseHuggingFaceTB/SmolVLM-256M-Instructrobot_offline_edgerank 8 / ≤0.5B recommended; throughput edge×robot_offline--action-scheme bins; pull presetssmolvlm-256m/smollm2-135mpython scripts/robot_offline_demo.pyTest plan
pytest tests/test_robot_offline.py(+ related) greenpython scripts/robot_offline_demo.py --steps 2onfake://