2nd/10_Wiki/Topics/AI_and_ML/Proprioception.md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	tech_stack
wiki-2026-0508-proprioception	Proprioception	10_Wiki/Topics	verified	self	Body Awareness Kinesthesia Sixth Sense	none	A	0.9	applied	neuroscience sensorimotor robotics embodied-ai		2026-05-10	pending	language framework Python ROS2/MuJoCo

Proprioception

매 한 줄

"매 body 의 self-knowledge — joint angle, muscle tension, limb position". 1906 Sherrington 이 명명. 매 sixth sense — vision/hearing 없어도 매 finger 의 코 의 touch 가능. 매 2026 robotics + embodied AI 에서 매 proprioceptive observation 의 핵심 input.

매 핵심

매 receptors (생물)

• Muscle spindles: 매 stretch + velocity 감지. Ia/II afferent.
• Golgi tendon organs: 매 muscle force 감지. Ib afferent.
• Joint receptors: 매 angle, end-range.
• Cutaneous mechanoreceptors: 매 skin stretch — proprioception 의 augment.
• Vestibular system: 매 head orientation — 매 proprioception 과 fuse.

매 processing

• 매 spinal reflex (knee jerk) → cerebellum (timing) → parietal cortex (body schema).
• 매 sensorimotor integration: 매 efference copy + reafference 의 비교.
• 매 body schema: 매 dynamic internal model — 매 phantom limb 의 disruption.

매 응용

1. Robotics — 매 joint encoder + IMU = robot proprioception.
2. Embodied AI / RL — 매 proprioceptive observation vector.
3. Rehabilitation — 매 stroke / Parkinson's.
4. VR / haptics — 매 mismatch 의 motion sickness.
5. Sports / dance training.

💻 패턴

Robot proprioceptive state (MuJoCo)

import mujoco
import numpy as np

model = mujoco.MjModel.from_xml_path("humanoid.xml")
data = mujoco.MjData(model)

def proprio_obs(data):
    return np.concatenate([
        data.qpos[7:],          # joint positions (skip root)
        data.qvel[6:],          # joint velocities
        data.sensordata,        # IMU, touch, force
    ])

RL with proprioception only (no vision)

import gymnasium as gym
import torch.nn as nn

# 매 proprioceptive policy — 매 vision 없이 locomotion 가능
class ProprioPolicy(nn.Module):
    def __init__(self, obs_dim, act_dim):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(obs_dim, 256), nn.ELU(),
            nn.Linear(256, 256), nn.ELU(),
            nn.Linear(256, act_dim),
        )
    def forward(self, obs):
        return self.net(obs)

Sim-to-real proprioception (domain rand)

# 매 real robot 의 sensor noise 의 simulate
def add_proprio_noise(qpos, qvel):
    qpos = qpos + np.random.normal(0, 0.01, qpos.shape)  # encoder noise
    qvel = qvel + np.random.normal(0, 0.05, qvel.shape)
    # latency
    return delay_buffer.push_pop(np.concatenate([qpos, qvel]))

IMU sensor fusion (Madgwick)

import numpy as np

class Madgwick:
    def __init__(self, beta=0.1, dt=0.01):
        self.q = np.array([1.0, 0, 0, 0])
        self.beta = beta; self.dt = dt

    def update(self, gyro, accel):
        # 매 gyro integration + accel correction
        # (매 simplified — 매 full impl 의 quaternion math)
        q = self.q
        qdot = 0.5 * quat_mul(q, np.array([0, *gyro]))
        # accel-based gradient correction omitted for brevity
        self.q = (q + qdot * self.dt)
        self.q /= np.linalg.norm(self.q)
        return self.q

Body schema (forward model)

class ForwardModel(nn.Module):
    """예측: q_t, action_t → q_{t+1}, expected_proprio"""
    def __init__(self, q_dim, a_dim):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(q_dim + a_dim, 128), nn.ReLU(),
            nn.Linear(128, q_dim))
    def forward(self, q, a):
        return q + self.net(torch.cat([q, a], dim=-1))

# efference copy: prediction error = surprise signal

Wearable proprioception (rehab)

# IMU on each segment → joint angle estimate
def joint_angle(imu_proximal, imu_distal):
    q_p = imu_proximal.quaternion
    q_d = imu_distal.quaternion
    q_rel = quat_mul(quat_conj(q_p), q_d)
    return quat_to_euler(q_rel)

Embodied AI (humanoid robot, 2026)

# 매 Figure / Tesla Optimus / Unitree style
# 매 proprioception 27-dim + vision 의 multimodal policy
obs = {
    "proprio": robot.read_proprio(),   # 27-dim
    "image": camera.read(),            # 224x224x3
    "language": "pick up the red cup",
}
action = policy(obs)  # diffusion policy or transformer

매 결정 기준

상황	Approach
Locomotion RL	Proprio-only (fast, robust)
Manipulation	Proprio + vision
Outdoor terrain	Proprio + IMU + vision
Whole-body humanoid	Proprio + vision + language
Rehabilitation feedback	Proprio + EMG

기본값: Joint encoder + IMU + foot contact = baseline robot proprioception.

🔗 Graph

• 부모: Embodied-AI
• 변형: Kinesthesia

🤖 LLM 활용

언제: 매 robotics task design, 매 obs space spec, 매 sim-to-real noise model. 언제 X: 매 quantitative neuroscience claim — 매 paper citation 의 필수.

❌ 안티패턴

• Sim에서만 perfect proprio: 매 real 의 noise/latency/drift 가 policy 의 break. 매 domain randomize.
• Vision-only locomotion: 매 occlusion / dark 의 brittle. 매 proprio fallback 필수.
• No efference copy: 매 forward model 의 학습 의 unstable. 매 action 의 condition.
• Ignoring foot contact: 매 binary contact signal 의 매 locomotion policy 의 critical input.

🧪 검증 / 중복

• Verified (Sherrington 1906, Proske & Gandevia 2012, modern RL robotics literature).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — proprioception biology + robotics application.