More screenshots
Works with the AI tools you already use
Claude Code
Cursor
Codex CLI
GitHub Copilot
Gemini CLI+20 moreClaude CodeCursorCodex CLIGitHub CopilotGemini CLI
OpenClaw
Manus
Grok+17 agentsRobotics Skills: 18 Fundamentals-First Advisors
18 robotics advisors (ROS 2, arms, hands, legged, drones, marine, safety, sim, learning, and more) that start from fundamentals, verify options live, and let you choose.
Free
Robotics Skills: 18 Fundamentals-First Advisors
Example session with this skill installed
I built a 12-DOF quadruped with hobby RC servos (position control only). Should I use MPC or RL to get it trotting?
- Read your context and instructions
- Compiled the robotics
Verdict: neither, yet. With position-only hobby servos, MPC and RL both fail to make the robot trot. The bottleneck is actuation, not the controller. On this hardware, plan for slow, statically stable gaits. If trotting is the real goal, change the actuators first.
Why (the fundamentals)
- A legged robot is underactuated. You can't command the floating base directly. Base motion comes only from contact forces at the feet, bounded by friction cones and push-only contact.
- A trot is a dynamic gait. For much of each stride only two diagonal feet are down, so the center of mass leaves the support polygon. Balance then has to be held by regulating those contact forces in real time.
- Both methods need torque. Locomotion MPC over centroidal or single-rigid-body dynamics outputs ground-reaction forces, and a whole-body controller turns them into joint torques. RL policies that transfer to hardware also depend on compliant, well-modeled actuators. A hobby servo hides a stiff internal position loop behind a high gear ratio, so you can't command torque, can't measure it, and can't backdrive the joint. Neither method gets the output it needs.
- What position servos do well is statically stable gaits: a crawl that keeps three feet down and the center of mass inside the support polygon. Support-polygon (ZMP-style) reasoning is the right tool for that regime.
Your options
| Option | What you get | Cost |
|---|---|---|
| A. Static crawl gait on this hardware (recommended now) | Reliable slow walking. You still learn gait timing, leg IK, and state estimation | Lowest. Keep the hardware |
| B. Rebuild with quasi-direct-drive, torque-controlled actuators | Makes trotting, MPC, and RL real options | New legs and motor drivers. The open reference for backdriveable modular actuators is MIT Mini Cheetah |
| C. Buy a torque-controlled commercial quadruped | Trots out of the box through a vendor SDK, so your work starts at control and learning | Money instead of build time. One example is Unitree Go2 |
If you choose B or C: MPC or RL?
- Model-based: OCS2 (https://github.com/leggedrobotics/ocs2), or Pinocchio + Crocoddyl (https://github.com/stack-of-tasks/pinocchio, https://github.com/loco-3d/crocoddyl). Predictable and debuggable, but needs a decent dynamics model.
- Learned: Isaac Lab (https://github.com/isaac-sim/IsaacLab) with RSL-RL (https://leggedrobotics.github.io/rsl_rl/). Robust on rough terrain, but sim-to-real succeeds or fails on actuator modeling: motor dynamics, torque limits, latency, and gear friction.
Before anything walks
- Do state estimation first. Fuse the IMU with leg kinematics. Estimator drift looks exactly like controller failure.
- Use a gantry or harness for the first tests. The first fall is a hardware bill.
- Stage it: sim → gantry → flat ground → terrain.
Open questions that would change this answer
- Is trotting the actual goal, or is "walks reliably" enough? If it's the latter, option A is the whole answer.
- Are these PWM hobby servos or serial bus servos (Dynamixel- or Feetech-class)? Some bus servos expose current or load readings. That allows limited compliance, but it's still not torque control.
- What's the budget ceiling? That decides between rebuilding (B) and buying (C).
Connects securely to your tools. The creator never sees your data.
About this skill
Ask a general assistant "which IK solver should I use?" and you get a confident paragraph. It might name a library that was archived two years ago, quote a spec it half-remembers, and skip the textbook method that was the right answer for your 6-DOF arm. Robotics punishes that. The field is old enough to have deep fundamentals and fast enough that half the tooling you remember is stale.
These 18 skills fix both ends. Every answer starts from the classic method and why it works. Then it adds modern alternatives verified against live sources in the same conversation. Then it stops and lets you choose.
How each skill works
- Frame. It pins down one decision at a time, in dependency order. Broad requests get split up and started upstream.
- Fundamentals. It explains the classic method and why it's shaped that way, with terminology defined in plain language before the math.
- Modern scan. It runs a live search on every invocation. Each skill ships a dated snapshot of its field with a source link on every entry, but treats it as a starting point to re-verify, never as the answer.
- You choose. It offers 2-4 real options, always including the classic method, with one marked as recommended and the reason stated.
- Apply, then loop. Your choice goes onto a running decision stack and the next decision comes up. If a new choice contradicts an earlier one, the skill stops and says so.
It has three modes. Guided (the default) handles one decision per turn. Fast-forward takes the recommended option and stops only for irreversible, budget, or hardware calls. Audit checks your existing setup against the sequence.
The skills
Cross-cutting:
robotics-advisor: kinematics, dynamics, and control theory, grounded in Craig's Introduction to Robotics with page citationsros2-master: topics vs services vs actions, lifecycle nodes, QoS, executors, and ros2_controlrobot-perception: sensor choice, intrinsic and hand-eye calibration, 3D/6D perception, fusion, and time syncrobot-safety: risk assessment, ISO 10218 and TS 15066, performance levels, e-stop architecture, and the certification pathrobot-sim: simulator choice, contact fidelity, domain randomization, and whether sim results predict hardwarerobot-learning: whether to learn at all, then demonstrations, policy class, evaluation, and a safety wrapper
Platforms:
robot-arm: URDF → IK → hardware interface → MoveIt 2 → calibration → safetyrobot-hand: grippers to five-finger hands, covering grasp strategy, grip force, in-hand manipulation, and teleop retargetingrobot-mobile: SLAM vs prebuilt maps, localization, Nav2 planners and costmaps, docking, and multi-floorrobot-legged: quadrupeds and humanoids, covering balance criteria, gait planning, MPC vs RL, and loco-manipulationrobot-aerial: multirotor vs VTOL, PX4 vs ArduPilot, GPS-denied flight, endurance budget, and failsafesrobot-marine: ROV vs AUV vs USV, thruster allocation, underwater navigation, pressure, and corrosionrobot-soft: soft and continuum robots, covering actuation, modeling limits, soft sensing, and when compliance wins
Systems and domains:
robot-fleet: Open-RMF, traffic negotiation, shared lifts, doors, and chargers, and DDS discovery at scalerobot-swarm: decentralized control, consensus, and when a swarm actually beats a coordinatorrobot-field: agriculture, construction, and inspection outdoors, RTK GNSS failure modes, and hardeningrobot-medical: surgical, rehab, and assistive robots, RCM mechanisms, pHRI, and the regulatory pathrobotics-radar: a maintenance sweep that re-verifies the snapshots and adds coverage for new robot types
You never need to remember which skill to call. Describe your robot, and the right skill triggers and hands off to its neighbors as the problem moves.
Why you'd trust it
- Nothing is presented unchecked. Every snapshot entry carries a live source URL, and answers pass that link through instead of asserting the fact bare.
- It catches stale-memory traps. For example,
osrf/rmf_core, the Open-RMF repo most training data still cites, was archived in 2021 and development moved to theopen-rmforg. - It's measured honestly. We ran 14 blind-judged questions over 9 rounds. The latest round came out at parity (93% vs 93%): the skills cite about 20 points better and score about 25 points worse on a "fabrication" criterion that turned out to measure whether a judge without search could confirm a reference. All 29 cited paper IDs came from verified snapshots. The full write-up, including a retracted earlier headline, is in the repo's EVAL.md.
Open source (MIT): https://github.com/joey114132/claude-robotics-skills
How to install
Works the same in every agent - Claude, Cursor, Codex, Copilot and 20+ more.
- 1
Download the ZIP
Free skills download straight away. Paid skills unlock right after purchase.
- 2
Unzip into your skills folder
Every agent reads skills from one folder on your machine. Drop the unzipped folder in there.
- 3
Ask your agent to use it
Restart the agent if it was already running. It picks the skill up automatically - no config needed.
Skills folder by agent
Click the path to copy it. Create the folder if it does not exist yet.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill passes our trust checks below.
Security scanned
Passed our 8-point scan before listing
Fresh listing
Recently published to Agensi
Free forever
No account required to browse
Trust & safety
Security scanned
Verified clean 2 days ago
- Free to download with an account