Skip to content

Installation

RoboSmith targets Python 3.10 or newer and is developed against Python 3.11 and 3.12. The base package installs the CLI, Pydantic models, LiteLLM integration, configuration loading, logging, and structured output helpers.

Terminal window
git clone https://github.com/Shaswat2001/robosmith.git
cd robosmith
pip install -e .

For the most common end-to-end workflow, install simulation, training, and agent extras:

Terminal window
pip install -e ".[sim,train,agent]"
ExtraInstallsUse when
simmujoco, gymnasiumYou want Gymnasium and MuJoCo environments.
traintorch, stable-baselines3You want SB3-backed RL training.
agentlanggraph, langchain-core, langchain-communityYou want robosmith run or robosmith auto.
roboticsgymnasium-roboticsYou want Fetch, Shadow Hand, and Adroit-style environments.
maniskillmani-skillYou want ManiSkill adapters.
videoimageio, imageio-ffmpeg, moviepyYou want rollout videos in delivery artifacts.
hubhuggingface-hubYou want to inspect or push Hub artifacts.
devpytest, ruff, mypy, pre-commitYou are developing RoboSmith itself.
allAll declared extrasYou want the broadest local install.
Terminal window
pip install -e ".[all]"

RoboSmith loads .env.local and .env from the working directory. It auto-detects the provider from whichever key is present.

Terminal window
ANTHROPIC_API_KEY=sk-ant-...
OPENAI_API_KEY=sk-...
GEMINI_API_KEY=AIza...
GROQ_API_KEY=gsk_...
OPENROUTER_API_KEY=sk-or-...
S2_API_KEY=...

Provider auto-detection priority is:

  1. Anthropic
  2. OpenAI
  3. Gemini
  4. Groq
  5. OpenRouter

The integration commands can be useful without keys. inspect, diag, and gen wrapper --no-llm can run without calling an LLM.

Terminal window
robosmith version
robosmith deps
robosmith trainers
robosmith envs

robosmith deps reports installed and missing packages for environment adapters, training backends, video, HuggingFace Hub, and Semantic Scholar support.

A dry run confirms the CLI and configuration without starting training:

Terminal window
robosmith run \
--task "Train a HalfCheetah to run as fast as possible" \
--dry-run

Some robotics suites have install steps outside normal Python extras.

SuiteNotes
Isaac LabRequires Isaac Sim and the Isaac Lab installation flow. RoboSmith has an isaac_lab adapter and registry entries, but availability depends on your local Isaac setup.
LIBERORequires LIBERO’s own package and benchmark assets. RoboSmith keeps the adapter boundary separate from the rest of the pipeline.
ManiSkillDeclared as .[maniskill]; GPU and simulator requirements still follow ManiSkill’s own documentation.
SymptomLikely causeCheck
No available trainer supports algorithmTraining extra is missing or backend dependencies are unavailable.Run robosmith trainers and install .[train].
Framework requires ...Environment adapter dependency is missing.Run robosmith deps and install the relevant suite.
LLM calls fail immediatelyNo supported API key was loaded.Check .env.local in the command working directory.
Videos are missingDelivery ran without video extras or environment render support.Install .[video] and choose render-capable environments.