v0.2.0

RoboSmith 0.2.0

RoboSmith 0.2.0 expands the project from a natural-language-to-policy prototype into a broader agentic robotics toolchain. This release adds the full LangGraph training workflow, integration tooling for existing robotics artifacts, trajectory diagnostics, adapter generation, run management, multi-provider LLM support, and a refreshed documentation site.

Highlights

• Added the checkpointed robosmith run LangGraph pipeline:
  intake -> scout -> env_synthesis -> inspect_env -> reward_design -> training -> evaluation -> delivery.
• Added resumable pipeline runs with per-node checkpoints, persisted run_state.json, artifacts directories, progress logs, and robosmith resume.
• Added literature-aware reward design with Semantic Scholar, arXiv, and combined scout backends.
• Added multi-provider LLM resolution through LiteLLM, .env.local loading, and support for Anthropic, OpenAI, Gemini, Groq, and OpenRouter.
• Added integration tooling for inspecting policies, datasets, environments, robots, compatibility mismatches, and generated adapters.
• Added trajectory diagnostics and rollout comparison commands.
• Added a run-management CLI for listing, inspecting, comparing, and cleaning past pipeline runs.
• Added a new Astro Starlight documentation site with guides, workflows, architecture, configuration, CLI, Python API, environment, trainer, and extension references.

New CLI Surface

• robosmith inspect dataset
• robosmith inspect env
• robosmith inspect policy
• robosmith inspect robot
• robosmith inspect compat
• robosmith diag trajectory
• robosmith diag compare
• robosmith gen wrapper
• robosmith auto integrate
• robosmith runs list
• robosmith runs inspect
• robosmith runs compare
• robosmith runs clean
• robosmith resume
• robosmith deps
• robosmith trainers
• robosmith envs

Training And Environment Improvements

• Added an environment registry with 30 entries across Gymnasium, MuJoCo, Gymnasium-Robotics, and Isaac Lab-style tasks.
• Added adapter infrastructure for Gymnasium, ManiSkill, Isaac Lab, LIBERO, and custom MJCF/URDF-style environments.
• Added trainer registry and backend interfaces for SB3, CleanRL, rl_games, imitation learning, and offline RL.
• Added automatic trainer selection by algorithm/backend availability.
• Improved support for dict observations, including robotics environments such as HandReach-v3.
• Improved environment synthesis and matching from natural-language task descriptions.
• Added video/report delivery artifacts for completed runs.

Agentic Workflow Improvements

• Added BaseAgent infrastructure for LiteLLM calls, retries, JSON parsing, and usage tracking.
• Added reward-generation and reward-evolution agents.
• Added a decision agent for post-evaluation routing: accept, refine reward, switch algorithm, or adjust environment.
• Added feedback loops so failed evaluations can feed training reflections back into reward design.
• Added stage skipping for selected stages such as scout, intake, and delivery.
• Added clearer terminal progress output and log-file capture.

Integration Tooling

• Added dataset inspection for LeRobot-style datasets, including schema, quality checks, sample frames, action/state dimensions, cameras, episodes, and tasks.
• Added policy inspection for model architecture, expected inputs/outputs, action heads, configs, and requirements.
• Added environment inspection for observation/action spaces, render modes, sample steps, and observation documentation.
• Added robot inspection for URDF/MJCF-style robot descriptions.
• Added compatibility reports across policies, datasets, and environments.
• Added template-based adapter generation for camera remapping, action dimension adaptation, image resizing stubs, and normalization stubs.
• Added auto integrate, a LangGraph workflow that chains inspection, compatibility checking, wrapper generation, and final reporting.

Diagnostics

• Added trajectory analysis for HDF5, LeRobot dataset directories, and Hub-style identifiers.
• Added rollout metrics including success rate, episode length, reward stats, action statistics, clipping rate, and failure clusters.
• Added side-by-side trajectory comparison with success-rate deltas, episode-length deltas, action divergence, and biggest degradation summaries.

Documentation

• Added the active docs site under robosmith-docs/ using Astro Starlight.
• Added installation, quickstart, training pipeline, integration tooling, runs, architecture, configuration, CLI, Python API, reward design, training backend, environment, and extension docs.
• Refreshed README examples around the two main workflows: train from scratch and integrate existing robotics work.
• Updated project branding, logo assets, and documentation styling.

Fixes And Polish

• Improved CLI organization and Rich output formatting.
• Improved robosmith envs filtering and empty-result messages.
• Improved logging noise suppression for LLM/http clients.
• Added clearer optional dependency checks through robosmith deps.
• Improved packaging extras for simulation, training, agent workflows, robotics environments, video, Hub integration, and development.
• Added focused test coverage for agent routing, inspection, diagnostics, config, registry, training, evaluation, wrapper generation, reward design, delivery, and environment synthesis.

Notes

• RoboSmith remains alpha software.
• Heavy robotics dependencies are optional and loaded lazily where possible.
• Some backends and environment suites depend on external installation steps, especially Isaac Lab, LIBERO, ManiSkill, and rl_games.
• inspect, diag, and gen wrapper --no-llm can be used without an LLM API key.
• robosmith run requires a supported LLM provider key or an explicit model configuration.