dora-rs vs SkillClaw

dora-rs is a Rust-native robotics middleware framework built around a declarative dataflow architecture — pipelines are defined as directed graphs in YAML, and nodes communicate through typed, Apache Arrow-formatted messages with zero serialization overhead. The project benchmarks at 10-17x faster than ROS2 Python, using zero-copy shared memory IPC for messages over 4KB and Zenoh for cross-machine pub-sub with 35% lower latency on large payloads than conventional messaging. What makes dora stand out from the crowded robotics-middleware space is that it was built to be agent-native from day one. The entire codebase is maintained through autonomous AI agents — a kind of recursive proof-of-concept for agentic software development. Nodes can be written in Rust, Python, C, or C++, hot reload is supported for Python operators, and built-in OpenTelemetry tracing is included without extra config. The framework is Apache 2.0 licensed and gaining traction with robotics researchers building real-time systems, self-driving stacks, and embodied AI demos. With 3.6k GitHub stars and an active Discord, it's early but credible as an alternative to ROS2 for teams who care about performance and composability.

SkillClaw is a research framework from Alibaba's AMAP-ML team that enables collective skill evolution for LLM agent systems deployed at scale. The core idea: instead of each user's agent interactions existing in isolation, SkillClaw aggregates anonymized skill-improvement signals across all users to continuously refine a shared library of reusable agent skills — without requiring centralized fine-tuning. The framework introduces a three-component architecture: a Skill Extractor that identifies and catalogs atomic capabilities from interactions, a Skill Evolver that proposes improvements based on aggregate feedback, and a Skill Selector that routes tasks to the best-available skill version per user context. Published on April 9 and hitting #1 on Hugging Face trending papers this week with 277 upvotes, the paper reports significant improvements over per-user baselines on complex multi-step agentic tasks. This matters especially for production agent deployments where cold-start problems are severe — a new user's agent immediately benefits from millions of prior interactions. It's a fundamentally different model of agent improvement than either fine-tuning (expensive, periodic) or RAG (retrieval-only, no learning).