feat(agents): Add hierarchical language support for VLA training#264
Draft
yuecideng wants to merge 1 commit into
Draft
feat(agents): Add hierarchical language support for VLA training#264yuecideng wants to merge 1 commit into
yuecideng wants to merge 1 commit into
Conversation
Add comprehensive language support to Online Data Streaming (ODS) for Vision-Language-Action (VLA) model training. The implementation provides: - Hierarchical language structure (task/subtask/primitive levels) - Multiple language sources (file, env, template, LLM) - Flexible storage modes (tokens, embeddings, hybrid) - LanguageManager for tokenization and data management - Integration with ODS shared memory buffer New files: - embodichain/lab/gym/envs/managers/language.py: LanguageManager, configs - embodichain/lab/gym/envs/managers/language_provider.py: Language providers - configs/language/: Example configurations and documentation - tests/agents/test_language_support.py: Test suite Modified files: - embodichain/agents/engine/data.py: Add language_cfg to OnlineDataEngine - embodichain/lab/gym/envs/embodied_env.py: Integrate LanguageManager - embodichain/lab/gym/utils/gym_utils.py: Extend buffer initialization - embodichain/lab/gym/envs/managers/__init__.py: Export language classes This enables VLA models to learn from multi-scale language representations similar to human task understanding. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
yuecideng
added
enhancement
Contributor
There was a problem hiding this comment.
Pull request overview
This PR introduces hierarchical language support for Vision-Language-Action (VLA) training by extending the Online Data Streaming (ODS) rollout buffer schema and integrating language generation/tokenization into EmbodiedEnv.
Changes:
- Adds
LanguageCfg,LanguageManager, and hierarchical language data structures. - Adds language providers (file/env/template/LLM) and wires language collection into environment episode initialization.
- Extends rollout buffer initialization to optionally allocate a
languageTensorDict and adds a test suite + example configs/docs.
Reviewed changes
Copilot reviewed 10 out of 10 changed files in this pull request and generated 20 comments.
Show a summary per file
| File | Description |
|---|---|
embodichain/lab/gym/utils/gym_utils.py |
Adds _init_language_buffer() and extends init_rollout_buffer_from_config() to allocate language fields. |
embodichain/lab/gym/envs/managers/language.py |
Implements language config + tokenization + hierarchical data formatting for buffers. |
embodichain/lab/gym/envs/managers/language_provider.py |
Adds language sources/providers (file/env/template/LLM). |
embodichain/lab/gym/envs/embodied_env.py |
Integrates language manager/provider and writes language into the rollout buffer at episode init. |
embodichain/agents/engine/data.py |
Plumbs language_cfg through ODS config into env config and buffer allocation. |
embodichain/lab/gym/envs/managers/__init__.py |
Re-exports new language modules. |
tests/agents/test_language_support.py |
Adds unit tests for language data structures/buffer init/providers. |
configs/language/README.md |
Documents configuration and buffer layout. |
configs/language/tasks_example.yaml |
Adds example hierarchical task descriptions. |
configs/language/usage_example.py |
Adds end-to-end usage examples. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Comment on lines
+995
to
+1000
| # Token IDs: [batch_size, max_episode_steps, max_instructions, max_tokens] | ||
| language_desc[f"{level_key}_tokens"] = torch.zeros( | ||
| (batch_size, max_episode_steps, max_instructions, max_tokens), | ||
| dtype=torch.int64, | ||
| device=device, | ||
| ) |
Comment on lines
367
to
371
| if self.cfg.init_rollout_buffer: | ||
| # Determine if we need to initialize language fields | ||
| language_cfg = self.cfg.language if self.cfg.language else None | ||
| self.rollout_buffer = init_rollout_buffer_from_gym_space( | ||
| obs_space=self.observation_space, |
Comment on lines
+776
to
+782
| # Write instruction count | ||
| count = buffer_format.get(f"{level_key}_count", torch.tensor([0])) | ||
| level_idx = {"task": 0, "subtask": 1, "primitive": 2}[level] | ||
| self.rollout_buffer["language"]["instruction_counts"][ | ||
| env_ids, :, level_idx | ||
| ] = count.item() | ||
|
|
Comment on lines
+390
to
+392
| # Stack instructions | ||
| result[f"{level_key}_tokens"] = torch.stack(padded_tokens) | ||
| result[f"{level_key}_attention_mask"] = torch.stack(padded_masks) |
Comment on lines
+374
to
+385
| # Empty instruction | ||
| tokens = torch.full( | ||
| (cfg.max_tokens,), | ||
| cfg.pad_token_id, | ||
| dtype=torch.int64, | ||
| device="cpu", | ||
| ) | ||
| mask = torch.zeros( | ||
| (cfg.max_tokens,), | ||
| dtype=torch.int64, | ||
| device="cpu", | ||
| ) |
Comment on lines
+633
to
+639
| # Would need LanguageManager to tokenize - return placeholder | ||
| return HierarchicalLanguageData( | ||
| task_level=[], # Would be populated with LanguageData objects | ||
| subtask_level=[], | ||
| primitive_level=[], | ||
| change_points=change_points, | ||
| ) |
| tokens, mask = self.tokenize(text) | ||
| return LanguageData(tokens=tokens, attention_mask=mask, raw_text=text) | ||
|
|
||
| temp_mgr = _TempManager(self.cfg) |
Comment on lines
+1256
to
+1258
| log_info( | ||
| f"[init_rollout_buffer_from_config] Language buffer added with hierarchy levels: {language_cfg.get('hierarchy_levels', ['task', 'subtask', 'primitive'])}" | ||
| ) |
Comment on lines
+345
to
+349
| self.language_manager = LanguageManager(language_cfg, self) | ||
| log_info( | ||
| f"[EmbodiedEnv] LanguageManager initialized with source={language_source}, " | ||
| f"mode={language_cfg.mode}, hierarchy={language_cfg.hierarchy_levels}" | ||
| ) |
Comment on lines
+795
to
+798
| # Write hierarchy depth | ||
| hierarchy_depth = language_data.hierarchy_depth | ||
| self.rollout_buffer["language"]["hierarchy_depth"][env_ids, :] = hierarchy_depth | ||
|
|
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
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.
Description
This PR adds comprehensive language support to Online Data Streaming (ODS) for Vision-Language-Action (VLA) model training. The implementation enables VLA models to learn from multi-scale language representations similar to human task understanding.
Key Features
Hierarchical Language Structure: Organizes instructions at three abstraction levels:
Multiple Language Sources:
Flexible Storage: Supports tokens, embeddings, or hybrid storage modes
LanguageManager: Handles tokenization and language data management with support for:
Changes
New files:
embodichain/lab/gym/envs/managers/language.py- LanguageManager, configs, and data structuresembodichain/lab/gym/envs/managers/language_provider.py- Language providers for different sourcesconfigs/language/- Example configurations, documentation, and usage examplestests/agents/test_language_support.py- Test suite (7 passed, 4 skipped due to optional dependencies)Modified files:
embodichain/agents/engine/data.py- Addedlanguage_cfgtoOnlineDataEngineCfgand buffer creationembodichain/lab/gym/envs/embodied_env.py- Integrated LanguageManager and language data writingembodichain/lab/gym/utils/gym_utils.py- Extendedinit_rollout_buffer_from_configto allocate language fieldsembodichain/lab/gym/envs/managers/__init__.py- Exported new language classesBuffer Structure
When language support is enabled, the rollout buffer includes:
{level}_tokens: Token IDs for each hierarchy level{level}_attention_mask: Attention masks for padding{level}_count: Number of instructions per levelinstruction_counts: Counts across all levelschange_points: Timesteps where language changeshierarchy_depth: Current depth of hierarchy (1-3)instruction_types: Instruction type IDsUsage Example
Type of change
Screenshots
N/A
Checklist
black .command to format the code base.🤖 Generated with Claude Code