MohammedSuleiman

Real three leg balance

Real hardwareClassical controlFront left leg

I wrote the balance controller from scratch. It shifts the robot's center of mass into the triangle formed by the three stance feet, then lifts the front left foot clear of the ground.

The lift was only one step. Jacobian based inverse kinematics converted foot targets into joint targets, joint level control moved the leg, and safety gates checked every phase. A failed check sent the robot to a stable recovery pose instead of letting the sequence continue.

• 01Center of mass shift — moved the torso before any foot left the ground.
• 02Jacobian IK — converted foot-space targets into hip, knee, and abduction angles.
• 03Safety gates — checked limits and stability between phases, with a recovery pose always armed.
• 04Staged testing — validated each phase alone before enabling the next.

Learned rear leg balance

Isaac LabReinforcement learningDomain randomization

For rear leg balance, I used reinforcement learning instead of a hand written controller.

I trained the policy in Isaac Lab with 2,400 robots running in parallel, and it converged in roughly six hours. Each simulated Spot had a randomized 2.4–3.0 kg payload and randomized floor friction, so the policy learned to recover under changing conditions instead of memorizing one model. It held the two leg stance in simulation; hardware deployment is planned for spring.

• ENV2,400 parallel environments · A5000 / A100
• WALL≈6 hours to a working policy
• RANDPayload and floor friction randomized for each robot every episode
• GOALSustained rear leg balance under unmodeled load
• STATEWorking in simulation · hardware deployment planned for spring

Repairing the hardware

3D printingSame day repair

When a motor mount broke, I found an existing printable model and made the replacement in the lab. The Crazyflie returned to testing the same day instead of waiting for a parts shipment.

The viewer above renders the actual STL used for the replacement. Drag the live mesh to inspect it.

Isaac Sim on FrostByte

LinuxSLURMEnrootA5000 / A100

Isaac Sim normally expects a workstation and a display. FrostByte provided neither, so with the help of IT, I containerized the runtime with Enroot, requested GPUs through SLURM, and built a headless rendering path. That let us use Amherst's A5000 and A100 GPUs instead of paying for cloud compute.

I documented the container, cluster commands, and launch process after the environment worked.

One managed service

One monthly price covers design, development, hosting, maintenance, and later updates. We usually launch in days, and we keep revising until the client approves the site.

• DESIGNCustom, mobile first layouts built around each client's business.
• HOSTManaged hosting, domains, uptime, and routine maintenance included.
• SPEEDSub second loads through tuned Core Web Vitals and WebP images.
• REACHSitemaps, metadata, analytics, and Google Business Profile setup.

Built for small businesses

We work with restaurants, churches, and local service businesses that need a professional site but do not have someone to manage it. Their owners want customers to find accurate information without handling domains, hosting, or updates.

The client approves the design and sends us changes when needed. We handle the technical work after that, and most new projects come through referrals from existing clients.

• SCOPEDesign, development, hosting, and maintenance through one team.
• PACESites launch in days, with revisions through final client approval.
• AFTERClients send new hours, menus, or text changes; we update the site.