convez.ai
AgenteAutomaçãoAvançado
something06030

RobotMem — Memória de Experiências para Robôs

something06030/robotmem

Armazena e recupera experimentos robóticos para otimizar sucesso de tarefas

Ver no GitHub
BomPluginScore · de 0 a 100

Agent que armazena e recupera dados de experimentos robóticos (parâmetros, trajetórias, resultados) usando busca híbrida (BM25 + vetores) com filtros estruturados e busca espacial. Melhora a taxa de sucesso de tarefas robóticas ao aprender com experiências passadas sem reiniciar do zero a cada tentativa.

Instalação

$npx convez-ai something06030/robotmem

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
3
Forks
0
Acompanhando
1
Saúde
57%
Último commit
há 1 mês
Versão
—
Issues abertas
0
Contribuidores
2

Casos de uso

  • Otimização de parâmetros em robótica (força de preensão, trajetórias)
  • Aprendizado por reforço com memória estruturada de episódios
  • Consolidação de experiências e recuperação proativa de contextos relevantes

Veredito

Solução especializada para robótica que combina armazenamento estruturado com retrieval inteligente (filtros JSON + busca espacial). Implementação sólida com ganho comprovável (+25% em sucesso), ideal para pesquisa e aplicações industriais que precisam de aprendizado com episódios anteriores.

README

A tradução para o português deste README ainda está na fila.

中文版

robotmem — Let Robots Learn from Experience

Your robot ran 1000 experiments, starting from scratch every time. robotmem stores episode experiences — parameters, trajectories, outcomes — and retrieves the most relevant ones to guide future decisions.

FetchPush experiment: +25% success rate improvement (42% → 67%), CPU-only, reproducible in 5 minutes.

Quick Start

pip install robotmem
from robotmem import learn, recall, save_perception, start_session, end_session

# Start an episode
session = start_session(context='{"robot_id": "arm-01", "task": "push"}')

# Record experience
learn(
    insight="grip_force=12.5N yields highest grasp success rate",
    context='{"params": {"grip_force": {"value": 12.5, "unit": "N"}}, "task": {"success": true}}'
)

# Retrieve experiences (structured filtering + spatial nearest-neighbor)
memories = recall(
    query="grip force parameters",
    context_filter='{"task.success": true}',
    spatial_sort='{"field": "spatial.position", "target": [1.3, 0.7, 0.42]}'
)

# Store perception data
save_perception(
    description="Grasp trajectory: 30 steps, success",
    perception_type="procedural",
    data='{"sampled_actions": [[0.1, -0.3, 0.05, 0.8], ...]}'
)

# End episode (auto-consolidation + proactive recall)
end_session(session_id=session["session_id"])

7 APIs

APIPurpose
learnRecord physical experiences (parameters / strategies / lessons)
recallRetrieve experiences — BM25 + vector hybrid search with context_filter and spatial_sort
save_perceptionStore perception / trajectory / force data (visual / tactile / proprioceptive / auditory / procedural)
forgetDelete incorrect memories
updateCorrect memory content
start_sessionBegin an episode
end_sessionEnd an episode (auto-consolidation + proactive recall)

Key Features

Structured Experience Retrieval

Not just vector search — robotmem understands the structure of robot experiences:

# Retrieve only successful experiences
recall(query="push to target", context_filter='{"task.success": true}')

# Find spatially nearest scenarios
recall(query="grasp object", spatial_sort='{"field": "spatial.object_position", "target": [1.3, 0.7, 0.42]}')

# Combine: success + distance < 0.05m
recall(
    query="push",
    context_filter='{"task.success": true, "params.final_distance.value": {"$lt": 0.05}}'
)

Context JSON — 4 Sections

{
    "params":  {"grip_force": {"value": 12.5, "unit": "N", "type": "scalar"}},
    "spatial": {"object_position": [1.3, 0.7, 0.42], "target_position": [1.25, 0.6, 0.42]},
    "robot":   {"id": "fetch-001", "type": "Fetch", "dof": 7},
    "task":    {"name": "push_to_target", "success": true, "steps": 38}
}

Each recalled memory automatically extracts params / spatial / robot / task as top-level fields.

Memory Consolidation + Proactive Recall

end_session automatically triggers:

  • Consolidation: Merges similar memories with Jaccard similarity > 0.50 (protects constraint / postmortem / high-confidence entries)
  • Proactive Recall: Returns historically relevant memories for the next episode

FetchPush Demo

cd examples/fetch_push
pip install gymnasium-robotics
PYTHONPATH=../../src python demo.py  # 90 episodes, ~2 min

Three-phase experiment: baseline → memory writing → memory utilization. Expected Phase C success rate 10-20% higher than Phase A.

Architecture

SQLite + FTS5 + vec0
├── BM25 full-text search (jieba CJK tokenizer)
├── Vector search (FastEmbed ONNX, CPU-only)
├── RRF fusion ranking
├── Structured filtering (context_filter)
└── Spatial nearest-neighbor sorting (spatial_sort)
  • CPU-only, no GPU required
  • Single-file database ~/.robotmem/memory.db
  • MCP Server (7 tools) or direct Python import
  • Web management UI: robotmem web

Comparison

FeatureMemoryVLA (Academic)Mem0 (Product)robotmem
Target usersSpecific VLA modelsText AIRobotic AI
Memory formatVectors (opaque)TextNatural language + perception + parameters
Structured filteringNoNoYes (context_filter)
Spatial retrievalNoNoYes (spatial_sort)
Physical parametersNoNoYes (params section)
InstallationCompile from paper codepip installpip install
DatabaseEmbeddedCloudLocal SQLite

License

Apache-2.0

Ver o original no GitHub

Bom para

Relacionados