K-Scale Robotics Skill Collection

K-Scale Robotics Skill Collection

Trit: 0 (ERGODIC - coordination/​infrastructure) Color: #5B8DEE (Sky Blue) URI: skill://kscale#5B8DEE

Overview

This skill indexes the K-Scale Labs robotics ecosystem - a comprehensive open-source stack for building, training, and deploying humanoid robots. The collection follows GF(3) triadic organization with kos-firmware (+1) as the primary generator and mujoco-scenes (0) as the coordinator.

Skill Inventory

┌────────────────────────────────────────────────────────────────────┐
│                    K-SCALE SKILL ECOSYSTEM                          │
├────────────────────────────────────────────────────────────────────┤
│                                                                     │
│  PLUS (+1) - Generation/​Construction                               │
│  ┌─────────────────────────────────────────────────────────────┐   │
│  │ kos-firmware        #79ED91  Robot firmware & gRPC services │   │
│  └─────────────────────────────────────────────────────────────┘   │
│                                                                     │
│  ERGODIC (0) - Coordination/​Infrastructure                         │
│  ┌─────────────────────────────────────────────────────────────┐   │
│  │ mujoco-scenes       #9FD875  Scene composition for MuJoCo   │   │
│  └─────────────────────────────────────────────────────────────┘   │
│                                                                     │
│  MINUS (-1) - Analysis/​Verification                                 │
│  ┌─────────────────────────────────────────────────────────────┐   │
│  │ ksim-rl             #3A2F9E  RL training for locomotion     │   │
│  │ evla-vla            #DBA51D  Vision-language-action model   │   │
│  │ urdf2mjcf           #4615B7  URDF to MJCF conversion        │   │
│  │ kbot-humanoid       #5B45C2  K-Bot robot specifications     │   │
│  │ zeroth-bot          #8CC136  3D-printed humanoid platform   │   │
│  │ kscale-actuator     #B9172E  Robstride motor control        │   │
│  │ entropy-sim2real    #E85B8E  Entropy-driven sim2real        │   │
│  └─────────────────────────────────────────────────────────────┘   │
│                                                                     │
└────────────────────────────────────────────────────────────────────┘

GF(3) Balance Analysis

Current State: UNBALANCED (trit sum = -6)

Balanced Triads

The primary balanced triad anchors the ecosystem:

ksim-rl (-1) ⊗ kos-firmware (+1) ⊗ mujoco-scenes (0) = 0 ✓

Pattern: Each MINUS skill can form a balanced triad by reusing the (kos-firmware, mujoco-scenes) pair:

evla-vla (-1)         ⊗ kos-firmware (+1) ⊗ mujoco-scenes (0) = 0 ✓
urdf2mjcf (-1)        ⊗ kos-firmware (+1) ⊗ mujoco-scenes (0) = 0 ✓
kbot-humanoid (-1)    ⊗ kos-firmware (+1) ⊗ mujoco-scenes (0) = 0 ✓
zeroth-bot (-1)       ⊗ kos-firmware (+1) ⊗ mujoco-scenes (0) = 0 ✓
kscale-actuator (-1)  ⊗ kos-firmware (+1) ⊗ mujoco-scenes (0) = 0 ✓
entropy-sim2real (-1) ⊗ kos-firmware (+1) ⊗ mujoco-scenes (0) = 0 ✓

Recommendations for Balance

To achieve independent balanced triads (no skill reuse), add PLUS (+1) skills:

1. kscale-deploy (+1): Deployment automation, OTA updates
2. kscale-onboard (+1): On-robot compute orchestration
3. kscale-teleop (+1): Teleoperation and remote control
4. kinfer-runtime (+1): On-device inference runtime

Architecture

                           ┌─────────────────────┐
                           │      TRAINING       │
                           │  ┌───────────────┐  │
                           │  │   ksim-rl     │  │
                           │  │  (PPO, AMP)   │  │
                           │  └───────────────┘  │
                           │         │           │
                           │         ▼           │
                           │  ┌───────────────┐  │
                           │  │ mujoco-scenes │  │
                           │  │ (environments)│  │
                           │  └───────────────┘  │
                           └─────────┬───────────┘
                                     │
              ┌──────────────────────┼──────────────────────┐
              │                      │                      │
              ▼                      ▼                      ▼
     ┌─────────────────┐    ┌─────────────────┐    ┌─────────────────┐
     │   MODELS        │    │   PERCEPTION    │    │   TRANSFER      │
     │ ┌─────────────┐ │    │ ┌─────────────┐ │    │ ┌─────────────┐ │
     │ │ urdf2mjcf   │ │    │ │   evla-vla  │ │    │ │entropy-sim2r│ │
     │ │ kbot-humanoi│ │    │ │  (VLA)      │ │    │ │  (domain    │ │
     │ │ zeroth-bot  │ │    │ └─────────────┘ │    │ │  random.)   │ │
     │ └─────────────┘ │    └─────────────────┘    │ └─────────────┘ │
     └─────────────────┘                           └─────────────────┘
              │                                             │
              └──────────────────────┬──────────────────────┘
                                     │
                                     ▼
                           ┌─────────────────────┐
                           │    DEPLOYMENT       │
                           │  ┌───────────────┐  │
                           │  │ kos-firmware  │  │
                           │  │   (+1)        │  │
                           │  └───────────────┘  │
                           │         │           │
                           │         ▼           │
                           │  ┌───────────────┐  │
                           │  │kscale-actuator│  │
                           │  │ (CAN motors)  │  │
                           │  └───────────────┘  │
                           └─────────────────────┘

Usage

Training Pipeline

# Full K-Scale training pipeline
from ksim import PPOTask
from ksim.robots.kbot import KBotConfig
from mujoco_scenes import SceneBuilder, Terrain

class KBotWalkingTask(PPOTask):
    robot = KBotConfig(model_path="kbot.mjcf")
    
    def build_scene(self):
        scene = SceneBuilder()
        scene.add_terrain(Terrain.FLAT)
        scene.add_random_obstacles(count=5)
        return scene.to_mjcf()
    
    # Entropy-driven domain randomization
    physics_randomizers = [
        StaticFrictionRandomizer(scale=0.5),
        MassMultiplicationRandomizer(scale=0.2),
    ]

# Train
task = KBotWalkingTask()
task.run_training(num_envs=4096)

Deployment Pipeline

from pykos import KosClient

async def deploy_policy():
    async with KosClient("kbot.local:50051") as client:
        # Load trained policy
        await client.policy.load("walking_v1.onnx")
        
        # Start control loop
        await client.policy.start()
        
        # Monitor
        while True:
            state = await client.actuator.get_actuators_state()
            print(f"Position: {state.positions}")

Key Contributors

• codekansas (Ben Bolte): Core architecture across ksim, kos
• budzianowski: EdgeVLA, dataset standardization
• nfreq: KOS PolicyService, calibration
• WT-MM: Visualization, integration
• b-vm: Randomizers, disturbances

Repository Links

Related Skills

• jaxlife-open-ended (+1): Open-ended evolution for behavior discovery
• ergodicity (0): Ergodic theory foundations
• wobble-dynamics (0): Perturbation response analysis
• stability (-1): Dynamical system stability analysis

References

@misc{kscale2024,
  title={K-Scale Labs Open Source Robotics Stack},
  author={K-Scale Labs},
  year={2024},
  url={https://github.com/​kscalelabs}
}