Skip to content

Repository files navigation

RoboSkills

Python 3.10+ Build Homepage MIT License

RoboSkills is a low-cost, minimal set of agent workflows and tools for robotics.

Two Sharpa robot hands holding a red string in a cat's cradle
Sharpa hands playing cat's cradle, reconstructed from a human video.

Workflows

Researched bimanual lifting: 0/2 initially, 2/2 after revision, 3/3 on fresh checks; Blender replay
Auto Research
Improve policies and reusable skills through repeated simulation experiments.

Setting: Auto-research; no direct privileged simulator state in the policy API.

Blender render of an LLM agent putting a wine bottle on a drawer in LIBERO
Agent as Policy
An LLM observes an episode and chooses actions turn by turn.

Setting: Zero/one-shot demonstrations; no auto-research.

Two YAM arms thread a rope through a ring in an agent-built simulation
Agentic Simulation
Solve simulation tasks from natural-language instructions.

Setting: Privileged simulator information available.

A DROID demonstration split on a diagonal with its simulation, a Panda taking a marker out of a mug
Real2Sim
Build interactive simulations from demonstrations.

Setting: Physics simulation included.

Studio dance video split on a diagonal with Leonard and Michelle repeating both dancers' motion
Motion Generation
One agent turns a text, image or video request into a motion video.

Setting: Kinematic motion and rendering; no physics.

Features

Low cost

With suitable tools, agents reach comparable performance on fewer, cheaper tokens.

Real2Sim · DexYCB potted meat can Motion Generation · Tennis serve
Cost $0.09. 11% of Codex. Cost $0.00. 0% of Claude Code.
A hand lifting a can beside a Sharpa hand lifting it in physics, split on a diagonal A tennis serve beside the same motion on a Mixamo character, split on a diagonal
Agent as Policy · RoboDojo stack blocks Auto Research · RoboDojo press by number
Cost $0.16 with DeepSeek. 7% of Claude Code. Both 5/5 on 5 seeds. Cost $0.14 with DeepSeek. 13% of Claude Code. Both 3/3 on 3 seeds.
Blender render of DeepSeek as the policy stacking blocks in RoboDojo, layout 0 Blender render of DeepSeek's frozen auto-research solution pressing buttons by number in RoboDojo, test layout 2

Minimal

We commit to keeping each workflow under 1,000 lines of code.

Tools

Tool Does Reference
RoboKit kinematics, motion planning, robot hands Code
GraspKit learned grasps Code
SAM 3 text → masks Paper · Code
SAM 3D image → object mesh Paper · Code
MoGe image → depth Paper · Code
GeoCalib image → gravity Paper · Code
HaWoR video → hand motion Paper · Code
WiLoR image → hands (MANO) Paper · Code
GVHMR video → human motion Paper · Code
Kimodo text → G1 or human motion Paper · Code
Blender rendering Site · Code
CoACD mesh → convex collision parts Paper · Code
MuJoCo physics, IK, contact feedback, rendering Paper · Code

Installation

Requires Python 3.10 or newer:

uv add roboskills      # or: pip install roboskills

The workflows run from a clone (git clone https://github.com/openrobokit/roboskills && cd roboskills && uv sync); each workflow's README goes on from there.

Quick links

Releases

Packages

Contributors

Languages