{"id":49464835,"url":"https://github.com/st0ff3l/vlm-pbrs-reinforcement-learning","last_synced_at":"2026-04-30T12:03:46.542Z","repository":{"id":351450233,"uuid":"1211044221","full_name":"St0ff3l/VLM-PBRS-Reinforcement-Learning","owner":"St0ff3l","description":null,"archived":false,"fork":false,"pushed_at":"2026-04-15T03:10:46.000Z","size":176,"stargazers_count":0,"open_issues_count":0,"forks_count":0,"subscribers_count":0,"default_branch":"main","last_synced_at":"2026-04-15T04:27:19.043Z","etag":null,"topics":[],"latest_commit_sha":null,"homepage":null,"language":"Jupyter Notebook","has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":null,"status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/St0ff3l.png","metadata":{"files":{"readme":"readme.md","changelog":null,"contributing":null,"funding":null,"license":null,"code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":null,"support":null,"governance":null,"roadmap":null,"authors":null,"dei":null,"publiccode":null,"codemeta":null,"zenodo":null,"notice":null,"maintainers":null,"copyright":null,"agents":null,"dco":null,"cla":null}},"created_at":"2026-04-15T02:33:52.000Z","updated_at":"2026-04-15T03:10:50.000Z","dependencies_parsed_at":null,"dependency_job_id":null,"html_url":"https://github.com/St0ff3l/VLM-PBRS-Reinforcement-Learning","commit_stats":null,"previous_names":["st0ff3l/vlm-pbrs-reinforcement-learning"],"tags_count":null,"template":false,"template_full_name":null,"purl":"pkg:github/St0ff3l/VLM-PBRS-Reinforcement-Learning","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/St0ff3l%2FVLM-PBRS-Reinforcement-Learning","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/St0ff3l%2FVLM-PBRS-Reinforcement-Learning/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/St0ff3l%2FVLM-PBRS-Reinforcement-Learning/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/St0ff3l%2FVLM-PBRS-Reinforcement-Learning/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/St0ff3l","download_url":"https://codeload.github.com/St0ff3l/VLM-PBRS-Reinforcement-Learning/tar.gz/refs/heads/main","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/St0ff3l%2FVLM-PBRS-Reinforcement-Learning/sbom","scorecard":null,"host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":286080680,"owners_count":32463911,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2026-04-29T22:27:22.272Z","status":"online","status_checked_at":"2026-04-30T02:00:05.929Z","response_time":57,"last_error":null,"robots_txt_status":"success","robots_txt_updated_at":"2025-07-24T06:49:26.215Z","robots_txt_url":"https://github.com/robots.txt","online":true,"can_crawl_api":true,"host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":[],"created_at":"2026-04-30T12:03:36.655Z","updated_at":"2026-04-30T12:03:46.530Z","avatar_url":"https://github.com/St0ff3l.png","language":"Jupyter Notebook","funding_links":[],"categories":[],"sub_categories":[],"readme":"# VLM-PBRS: Hierarchical Hybrid Reinforcement Learning\n\nWelcome to the **VLM-PBRS** project. This repository introduces a state-of-the-art **Hierarchical Hybrid Reward Architecture** combining Vision-Language Models (VLMs) and classical Reinforcement Learning (RL) for Embodied AI tasks.\n\n## 🚀 The Architecture\n\nTraining RL agents purely on Vision-Language Models (VLMs) like LLaVA-7B suffers from severe spatial resolution limitations (VLMs cannot accurately regress continuous floating-point coordinates).\n\nWe solve this using a multi-tiered approach:\n1. **Low-Level Micro Kinematics**: Physics-based dense rewards (`kinematic_reward`) handle high-frequency motor control.\n2. **High-Level Macro VLM (PBRS)**: We use the robust Multiple Choice Question (MCQ) format to query the VLM for coarse spatial milestones (Left, Bottom, Right). The VLM acts as an intelligent supervisor.\n3. **RBF Visual Smoothing**: Asymmetric semantic VLM labels are smoothed into a fully differentiable 3D continuous potential field using a Visual Prototype Cache.\n4. **PBRS Integration**: VLM evaluations are integrated via Potential-Based Reward Shaping ($F = \\gamma\\Phi' - \\Phi$), ensuring mathematical immunity to reward hacking.\n\n## 🏃 Quick Start\n\nEnsure you have your Ollama server running locally with `llava:7b` (or swap to a cloud API in the client).\n\n```powershell\n# Run the complete Hybrid training loop with periodic evaluations\npython main.py --timesteps 300000 --run-eval\n```\n\n## 📂 Repository Structure\n\n- `main.py`: The single entry point for training the Hybrid Architecture.\n- `envs/visual_wrapper.py`: Contains the `AdaptiveVisualPBRS_Wrapper` with RBF smoothing and Hybrid Reward calculations.\n- `vlm/llava_client.py`: The MCQ parser forcing strict categoric classification from the VLM.\n- `archive_pure_vlm/`: Backups of the \"pure visual\" codebase configurations for future rigorous benchmarking (e.g., when massive API models replace local 7B models).\n- `archive_docs_and_old_tests/`: Old proposals and redundant test files.\n\n## 📜 Readings\n\nFor academic context, please review `PAPER_HYBRID_PROPOSAL.md` and `REVIEW_HYBRID_HIERARCHICAL_VLM.md`.\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fst0ff3l%2Fvlm-pbrs-reinforcement-learning","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Fst0ff3l%2Fvlm-pbrs-reinforcement-learning","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fst0ff3l%2Fvlm-pbrs-reinforcement-learning/lists"}