Skip to main content 홈 크리에이터 adu2021 skillxiv robofactory-embodied-agent-collaboration
robofactory-embodied-agent-collaboration Learn coordinated manipulation behaviors for multi-robot systems using compositional constraints that enforce safe and efficient collaboration. Generate training data through automated collection with task-specific constraint interfaces, then train imitation learning policies adaptable to varying difficulty levels.
설치로 이동 Skills Marketplace 커뮤니티가 만든 AI 스킬을 발견하고 탐색하세요.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
직접 명령은 검토 Prompt를 거치지 않습니다. 실행하기 전에 소스를 확인하세요.
npx skills add https://github.com/ADu2021/skillXiv --skill robofactory-embodied-agent-collaboration명령은 한 줄로 유지됩니다. 복사하기 전에 가로로 스크롤해 전체 내용을 확인하세요.
로컬 사본을 원하시나요? SkillsMP에서 현재 제공할 수 있는 파일을 다운로드하세요.
Zip 다운로드 다운로드 중... name robofactory-embodied-agent-collaboration title RoboFactory: Exploring Embodied Agent Collaboration with Compositional Constraints version 0.0.2 engine skillxiv-v0.0.2-claude-opus-4.6 license MIT url https://arxiv.org/abs/2503.16408 keywords ["Multi-Robot Systems","Embodied AI","Imitation Learning","Compositional Constraints","Manipulation Tasks"] description Learn coordinated manipulation behaviors for multi-robot systems using compositional constraints that enforce safe and efficient collaboration. Generate training data through automated collection with task-specific constraint interfaces, then train imitation learning policies adaptable to varying difficulty levels.
Core Concept
RoboFactory addresses the challenge of training multiple embodied agents to collaborate on complex manipulation tasks. The key innovation is compositional constraints—structured specifications that govern how agents should interact and coordinate. Rather than unstructured multi-agent learning, compositional constraints provide safety guardrails and efficiency guidance during both data collection and policy training. An automated data collection framework uses task-specific interfaces adapted to different constraint types, enabling safe generation of training trajectories. Imitation learning policies trained on this data learn coordinated behaviors that generalize across difficulty levels.
Architecture Overview
The system has three primary components:
Compositional Constraint Specification : Formal definitions of coordination rules and safety boundaries for multi-agent systemsConstraint-Aware Data Collection : Automated framework using task-specific interaction mechanisms to generate training trajectories that respect constraintsImitation Learning Policies : Neural network policies trained from constraint-compliant demonstrations, adapted for varying task difficulty
The framework emphasizes both safety (constraints prevent dangerous collisions or invalid states) and efficiency (constraints guide agents toward coordinated solutions).
Implementation
The compositional constraint system defines valid multi-agent state spaces and transitions:
import numpy as np
from typing import List , Dict , Tuple
from abc import ABC, abstractmethod
class CompositionConstraint (ABC ):
"""Base class for compositional constraints in multi-robot systems."""
@abstractmethod
def is_valid (self, agent_states: Dict [str , np.ndarray] ) -> bool :
"""Check if current agent configuration satisfies constraint."""
pass
@abstractmethod
def
get_allowed_actions
self, agent_id: str , current_state: np.ndarray
"""Return mask of valid actions for agent given current state."""
pass
def
sample_valid_state
self
Dict
str
"""Sample a valid multi-agent configuration."""
pass
class
CollisionAvoidanceConstraint
CompositionConstraint
"""Prevents robots from colliding with each other or environment."""
def
__init__
self, min_distance: float = 0.1
self
def
is_valid
self, agent_states: Dict [str , np.ndarray]
bool
"""
Args:
agent_states: dict mapping agent_id -> position (3,)
Returns:
True if all pairwise distances exceed min_distance
"""
list
for
in
enumerate
for
in
1
3
3
if
self
return
False
return
True
def
get_allowed_actions
self, agent_id: str , current_state: np.ndarray
"""Return action mask preventing collision trajectories."""
7
bool
for
in
range
7
self
if
not
self
False
return
def
_simulate_action
self, state: np.ndarray, action_idx: int
"""Predict next state given action."""
0.01
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
3
return
def
_is_collision_free
self, position: np.ndarray
bool
"""Check if position is collision-free."""
return
True
class
CoordinationConstraint
CompositionConstraint
"""Enforces task-specific coordination patterns."""
def
__init__
self, pattern: str = "simultaneous"
"""
Args:
pattern: 'simultaneous', 'sequential', or 'interleaved'
"""
self
self
0
def
is_valid
self, agent_states: Dict [str , np.ndarray]
bool
"""Validate that agents follow coordination pattern."""
if
self
"simultaneous"
return
all
6
0.5
for
in
elif
self
"sequential"
sum
1
for
in
if
6
0.5
return
1
return
True
def
get_allowed_actions
self, agent_id: str , current_state: np.ndarray
"""Restrict actions based on coordination pattern."""
7
bool
if
self
"sequential"
6
0.5
if
not
1
False
return
The constraint-aware data collection framework generates safe training trajectories:
class ConstraintAwareDataCollector :
"""Generates imitation learning data respecting compositional constraints."""
def __init__ (self, constraints: List [CompositionConstraint], max_episode_length: int = 100 ):
self .constraints = constraints
self .max_episode_length = max_episode_length
def collect_episode (self, task: str , agent_ids: List [str ] ) -> Dict :
"""
Collects one episode of multi-agent trajectories.
Args:
task: task description (e.g., 'pick_and_place')
agent_ids: list of robot identifiers
Returns:
episode: dict with states, actions, rewards
"""
episode = {
'states' : {agent_id: [] for agent_id in agent_ids},
'actions' : {agent_id: [] for agent_id in agent_ids},
'rewards' : [],
'valid' : True
}
agent_states = self ._initialize_valid_state(agent_ids)
for step in range (self .max_episode_length):
for agent_id in agent_ids:
episode['states' ][agent_id].append(agent_states[agent_id].copy())
allowed_actions = {}
for agent_id in agent_ids:
masks = [
c.get_allowed_actions(agent_id, agent_states[agent_id])
for c in self .constraints
]
allowed_actions[agent_id] = np.logical_and.reduce(masks)
actions = {}
for agent_id in agent_ids:
valid_action_indices = np.where(allowed_actions[agent_id])[0 ]
if len (valid_action_indices) == 0 :
episode['valid' ] = False
break
action_idx = np.random.choice(valid_action_indices)
actions[agent_id] = action_idx
episode['actions' ][agent_id].append(action_idx)
if not episode['valid' ]:
break
agent_states = self ._step_environment(agent_states, actions)
for constraint in self .constraints:
if not constraint.is_valid(agent_states):
episode['valid' ] = False
break
if not episode['valid' ]:
break
reward = self ._compute_task_reward(agent_states, task)
episode['rewards' ].append(reward)
return episode
def _initialize_valid_state (self, agent_ids: List [str ] ) -> Dict [str , np.ndarray]:
"""Initialize multi-agent state satisfying all constraints."""
agent_states = {}
for i, agent_id in enumerate (agent_ids):
state = np.zeros(7 )
state[:3 ] = np.array([i * 0.2 , 0 , 0 ])
agent_states[agent_id] = state
return agent_states
def _step_environment (self, agent_states: Dict [str , np.ndarray], actions: Dict [str , int ] ) -> Dict [str , np.ndarray]:
"""Simulate environment step with multi-robot physics."""
next_states = {}
for agent_id, state in agent_states.items():
action_idx = actions[agent_id]
next_states[agent_id] = state.copy()
if action_idx < 6 :
directions = [
[0.01 , 0 , 0 ], [-0.01 , 0 , 0 ],
[0 , 0.01 , 0 ], [0 , -0.01 , 0 ],
[0 , 0 , 0.01 ], [0 , 0 , -0.01 ]
]
next_states[agent_id][:3 ] += directions[action_idx]
return next_states
def _compute_task_reward (self, agent_states: Dict [str , np.ndarray], task: str ) -> float :
"""Compute task-specific reward signal."""
return 0.0 if task == "pick_and_place" else 0.0
Imitation learning policies are trained on constraint-compliant data:
class MultiAgentImitationPolicy (nn.Module):
"""Neural network policy for multi-robot imitation learning."""
def __init__ (self, state_dim: int = 7 , action_dim: int = 7 , hidden_dim: int = 256 ):
super ().__init__()
self .state_encoder = nn.Sequential(
nn.Linear(state_dim, hidden_dim),
nn.ReLU(),
nn.Linear(hidden_dim, hidden_dim)
)
self .action_head = nn.Linear(hidden_dim, action_dim)
self .confidence_head = nn.Linear(hidden_dim, 1 )
def forward (self, state: torch.Tensor ) -> Tuple [torch.Tensor, torch.Tensor]:
"""
Args:
state: (B, 7) agent state
Returns:
action_logits: (B, 7) action distribution
confidence: (B, 1) policy confidence [0, 1]
"""
encoded = self .state_encoder(state)
action_logits = self .action_head(encoded)
confidence = torch.sigmoid(self .confidence_head(encoded))
return action_logits, confidence
def train_imitation_policy (
episodes: List [Dict ],
policy: MultiAgentImitationPolicy,
num_epochs: int = 100 ,
learning_rate: float = 1e-3
):
"""
Train policy on constraint-compliant demonstrations.
Args:
episodes: list of collected episodes
policy: neural network policy
num_epochs: training epochs
learning_rate: optimizer learning rate
"""
optimizer = torch.optim.Adam(policy.parameters(), lr=learning_rate)
criterion = nn.CrossEntropyLoss()
for epoch in range (num_epochs):
total_loss = 0
for episode in episodes:
if not episode['valid' ]:
continue
agent_ids = list (episode['states' ].keys())
for agent_id in agent_ids:
states = torch.tensor(np.array(episode['states' ][agent_id]), dtype=torch.float32)
actions = torch.tensor(np.array(episode['actions' ][agent_id]), dtype=torch.long)
action_logits, confidence = policy(states)
loss = criterion(action_logits, actions)
optimizer.zero_grad()
loss.backward()
optimizer.step()
total_loss += loss.item()
if (epoch + 1 ) % 10 == 0 :
print (f"Epoch {epoch + 1 } , Loss: {total_loss / len (episodes):.4 f} " )
Practical Guidance
Training multi-robot systems for collaborative manipulation tasks
Scenarios where safety and coordination are critical
Tasks with clear compositional structure (collision avoidance, turn-taking, etc.)
Difficulty-varying domains (easy, medium, hard) requiring generalization
Imitation learning settings with access to expert demonstrations
Single-robot control (overhead not justified)
Fully decentralized systems with no coordination requirements
Real-time learning scenarios (constraint checking adds overhead)
Unstructured environments where compositional constraints don't apply
Reinforcement learning from scratch without demonstrations
min_distance: 0.05-0.2 meters for collision avoidancemax_episode_length: 100-500 steps depending on task complexityhidden_dim: 256-512 for policy networkslearning_rate: 1e-3 to 1e-4 for stable trainingConstraint combination strategy: AND (all constraints) vs. weighted sum
Constraints too restrictive, leaving few valid actions (training data bottleneck)
Constraints too loose, allowing unsafe or inefficient behaviors
Not verifying constraints are satisfiable before data collection
Overfitting to constraint interface; poor generalization to real robots
Ignoring task rewards in favor of purely constraint satisfaction
Performance Notes
Benchmark: RoboFactory embodied multi-agent manipulation
Difficulty levels: Evaluated across easy, medium, hard configurations
Imitation learning success rate: Varies by task and constraint complexity
Training time: Hours to days depending on episode count and model size
Generalization: Policies trained on medium difficulty generalize to hard difficulty
References
Multi-agent reinforcement learning and coordination
Imitation learning from demonstrations
Compositional AI and constraint-based systems
Robot manipulation and multi-robot coordination
Safety in embodied AI systems
