任务之外也在观察！智元发布终身具身智能框架LIFE 解锁机器人持续在线能力

Summary: AgiBot's LIFE, a lifelong embodied intelligence framework, lets robots stop forgetting after a task and continuously observe the environment during and between tasks, record what they see, and reuse it later.

1. Three components split responsibilities: RoboBeats comes in 0.8B, 2B, and 4B lightweight models and only judges whether a condition holds rather than generating long frame-by-frame text; plastic memory stores both raw images and structured JSON records, with images preserving scene details and records capturing time, location, people, and objects; EDM turns explored methods into tools that can be called on demand.

2. Results are concrete: on a single RTX 5090, processing a new frame takes about 75 ms, and detecting 16 conditions simultaneously takes about 77 ms; proactive perception accuracy rises from 21.2% to 64.0%; action-boundary detection F1 rises from 26.4 to 45.8; in a 24-hour real-robot test, LIFE answers 29.5/30 across six questions in 80.6 seconds on average, while the control group scores 28.5/30 but takes 554.7 seconds.

3. It is immediately usable: the paper, project page, code (AgibotTech/LIFE on GitHub), three model weights, and SSS-Bench evaluation resources are all open, so developers can reproduce it the same day.

Summary: For brands, the value lies in the technical narrative and open-source reputation: AgiBot frames lifelong embodied intelligence as a clear product direction with verifiable metrics and uses full open source to build industry influence.

1. Product R&D direction: the framework extends robot capability from single-task execution to cross-task accumulation; continuous perception, plastic memory, and capability reuse form the capability map for next-generation products and can serve as an R&D roadmap reference.

2. Brand marketing playbook: the paper, project page, GitHub code, three model weights, and evaluation benchmark are released together, with emphasis on same-day developer reproduction, a typical open-source-first approach to ecosystem and mindshare.

3. Communication assets: 75 ms on-device latency, 64.0% proactive perception accuracy, 29.5/30 real-robot Q&A, and roughly 45% token reduction are all directly usable competitive proofs.

4. User behavior observation: the article stresses that the environment keeps changing between tasks and that users want the machine to stay online and observe even without commands, suggesting product selling points should shift from obedient execution to long-term monitoring and experience accumulation.

Summary: For sellers, this signals new robot supply and new scenarios: open models and code lower the barrier to entry, while always-on perception opens demand for long-duration environmental monitoring.

1. Market opportunity: inspection, security, warehousing, and care scenarios that require all-day attention can be activated; plastic memory can turn a day's footage into searchable records and answer historical questions such as where an item was last placed.

2. Cost and barrier: the three lightweight models run on a single RTX 5090, producing a frame-level conclusion in about 75 ms and detecting 16 conditions simultaneously in about 77 ms, making deployment possible even for small and midsize sellers.

3. What to learn: EDM codifies a one-off solution into a tool mounted in a tool library; on the same set of queries not used to build the capability, success rate rises from 37.5% to 65.8%, meaning delivered capability can continue to appreciate.

4. Risk note: the article explicitly calls this a technical exploration and says multi-day online deployment and dynamic simulation are still needed to validate memory updates, decision-making, and operational reliability, so sellers should not promise mature commercial readiness to customers too early.

Summary: Factories can read this article through three lenses: edge deployment, on-site monitoring, and experience capture. It outlines a low-cost path to turn visual observation into reusable experience.

1. Production and design needs: the perception side only needs to judge whether conditions hold, such as whether a grasp is complete or an elevator door is open; increasing conditions from 1 to 16 raises latency only from 75 ms to 77 ms, making it suitable for parallel monitoring across workstations.

2. Business opportunity: 24-hour first-person video can be used offline to build memory and support historical Q&A, matching real needs such as production-line traceback, item ownership, and responsibility tracking; the article includes cases involving cup ownership and coat location.

3. Digitalization insight: memory has two layers, with raw images preserving fidelity and JSON structured fields making it easy to connect to existing information systems; EDM mounts retrieval and deduplication methods learned on site as tools, reducing repeated programming and offering a concrete approach to digitizing experience.

4. Deployment prerequisites: it runs on a single card, but long-term reliability still needs multi-day online deployment and dynamic simulation validation; small-scenario pilots are advisable first.

Summary: Service providers should check whether customer pain points and solutions align. LIFE targets the long-standing problem that robots reset after each task and cannot accumulate experience, and offers a quantifiable approach.

1. Industry trend: the scheduling unit in embodied intelligence is shifting from individual tasks to the intervals between tasks, with continuous perception, long-term memory, and capability reuse becoming standard framework features.

2. Customer pain point: during data collection it is hard to predict which details will matter later, and old records often lack fields. LIFE stores both raw images and JSON, so later users can revisit images to re-extract information such as door numbers and room numbers and add indexes, solving the problem of not recording something when needed and needing it later.

3. Solution: RoboBeats performs heartbeat-style monitoring to detect events, plastic memory handles retrieval, and EDM turns methods into tools; on SSS-Bench, Q&A accuracy rises from 40.9% to 60.5%, and on OVO-Bench the write-side token count drops by about 45%, keeping costs manageable.

4. Integration convenience: the paper, code, three model weights, and evaluation set are all open, making rapid proof-of-concept work and needs-based model selection easier.

Summary: Marketplace sellers can view this article as a model of using open source plus benchmarks to drive an ecosystem: AgiBot has built a developer entry point with a project page, code repository, model weights, and the SSS-Bench evaluation benchmark.

1. Platform playbook: the paper, project page, GitHub code, three RoboBeats-0.8B/2B/4B weights, and SSS-Bench resources are released together, with emphasis on same-day developer reproduction, clearly aiming to lower the onboarding barrier and broaden access.

2. Partner recruitment and operations: the three model sizes give customers with different compute budgets room to choose; six benchmark groups (on-device performance, proactive perception, action boundaries, memory effectiveness, retrospective Q&A, and capability reuse) create a common comparison standard for cross-ecosystem benchmarking and leaderboard-style operations.

3. Commercial needs and issues: the article notes that future work must combine long-term goals with autonomous action and be evaluated through multi-day online deployment and dynamic simulation, implying that platforms need to provide long-cycle test environments and real-robot video data services.

4. Hype avoidance: the article describes itself as a technical exploration and does not yet offer conclusions on large-scale commercialization; platforms should avoid overpromising in promotion and first make reproduction and evaluation capabilities solid.

Summary: Researchers can focus on three new directions: the paradigm shift from single tasks to cross-task operation, the engineering implementation of memory and capability reuse, and the open problems and evaluation resources that remain.

1. New paradigm: LIFE moves the scheduling unit from the task itself to the intervals between tasks and proposes heartbeat-style continuous perception without commands; on StreamingBench, proactive-output task accuracy rises from 21.2% to 64.0%, and on SPOT-Bench action-boundary F1 rises from 26.4 to 45.8.

2. New mechanisms: plastic memory uses dual storage of raw images plus JSON, enabling post hoc re-parsing and field and index supplementation; EDM does not update base model parameters and expands only at the program, tool-interface, and skill layers. In FindingDory, success rate on 597 queries not used to build the capability rises from 37.5% to 65.8%, an increase of 28.3 percentage points, or about 75% relative.

3. New questions and implications: the paper argues that future work needs to combine long-term goals with autonomous action and evaluate memory updates, decision-making, and operational reliability through multi-day online deployment and dynamic simulation; benchmarks such as SSS-Bench (171 videos, 618 questions) and the OVO-Bench retrospective subset are open for reproduction and comparison.

4. Key data: in a 24-hour real-robot experiment, each of six questions was tested three times; with the same Q&A model, LIFE scores 29.5/30 at an average of 80.6 seconds, while the Codex control group scores 28.5/30 at an average of 554.7 seconds, a striking efficiency difference.