{ "schema_version": 1, "builder_plan": { "scenario": "A support or data assistant repeats work across related customer requests.", "change": "Try the paper's memory change in a replay of past tasks, with separate memory per customer.", "baseline": "Your current memory pipeline plus a no-memory control using the same agent.", "steps": [ "Locate the paper's implementation and check its license, model support, data needs, and hardware requirements. Stop here if you cannot obtain the required artifacts.", "Save a small set of representative product tasks and expected outcomes. Keep evaluation tasks out of any training or tuning, and choose quality and cost limits before running.", "Run the baseline and the proposed change on identical inputs. Keep other settings fixed and repeat paired runs to expose variation.", "Save per-task outputs, failures, timings, and all usage costs. Review regressions before expanding the test." ], "measure": "Task success, stale memories, retrieval misses, response latency, and all memory maintenance costs.", "decision": "Keep the change only if it meets your preselected quality and latency limits and improves the metric you care about after setup and operating costs. Otherwise retain the baseline. This is an exploration plan, not a validated recommendation." }, "builder_review": null, "arxiv_id": "2609.16864", "title": "TEMPO: Learning Temporal Context for Dynamic Robot Manipulation", "summary": "TEMPO replaces single-observation inference in pretrained VLA policies with two added inputs: a motion summary from a frozen video foundation model and a compact proprioceptive history, requiring no backbone modification. The cost is the additional frozen video model inference at runtime and minimal training/deployment compute overhead, plus dependency on annotated temporal data (the authors release a 50k-frame benchmark). It can fail if the frozen video encoder's motion summaries do not transfer to the reader's camera setup or task dynamics, if proprioceptive history is unavailable or noisy on the target robot, and gains are demonstrated on only four dynamic manipulation tasks.", "rationale": "The method targets robot manipulation with pretrained VLA backbones and physical or simulated robot setups, which falls outside the reader's stated LLM-focused production scope, though the temporal-context principle is worth monitoring.", "technique": "Temporal context augmentation for vision-language-action models", "family": "memoria_e_contexto", "recommendation": "observar", "published": "2026-09-15", "as_of": "2026-10-18", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2609.16864", "independent_implementations": 1, "total_implementations": 2, "stars": 223, "citations": 0, "claims": [ { "claim_id": "claim-01", "statement": "TEMPO replaces single-observation inference in pretrained VLA policies with two added inputs: a motion summary from a frozen video foundation model and a compact proprioceptive history, requiring no backbone modification. The cost is the additional frozen video model inference at runtime and minimal training/deployment compute overhead, plus dependency on annotated temporal data (the authors release a 50k-frame benchmark). It can fail if the frozen video encoder's motion summaries do not transfer to the reader's camera setup or task dynamics, if proprioceptive history is unavailable or noisy on the target robot, and gains are demonstrated on only four dynamic manipulation tasks.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2609.16864", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "Improves Bottle Handover success rate from 44% to 74% across dynamic manipulation tasks and is the only evaluated method that resolves state aliasing.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2609.16864", "source_page": null, "source_excerpt": "" } ], "exposure_map": [ { "dimension": "quality", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "compute", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "latency", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "operations", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "compatibility", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "security", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "data_and_training", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "reproducibility", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" } ], "risks": [], "minimum_test": [], "open_questions": [], "independent_tests": [], "lab_reproductions": [], "report_available": false, "validation_tier": "unknown", "evidence_tier": "unknown", "infrastructure_basis": "unknown", "equations": [ { "anchor": "S4.E1", "label": "(1)", "section": "4 Temporal Encoding for Motion-aware Policy (TEMPO)", "role": "proposed_method", "latex": "\\mathbf{m}_{t}\\;=\\;\\phi(o_{t-W+1:t})\\;=\\;\\operatorname{CrossAttn}\\bigl(\\mathbf{f}_{t},\\;\\mathcal{F}_{t}\\bigr),", "mathml": "𝐦t=Ο•(otβˆ’W+1:t)=CrossAttn(𝐟t,β„±t),", "context": "TEMPOMOT summarizes recent visual dynamics by cross-attending the current frame against a rolling cache of past frames (Fig. 3). Let πŸΟ„ denote the patch features of observation oΟ„, extracted by ϕ’s per-frame encoder. Over a streaming window of length W, we maintain a cache β„±t={πŸΟ„}Ο„=tβˆ’W+1tβˆ’1 of features from the preceding Wβˆ’1 frames, populated incrementally so that each πŸΟ„ is computed exactly once and reused on subsequent steps at no additional cost. The motion code is then", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" }, { "anchor": "S4.E2", "label": "(2)", "section": "4 Temporal Encoding for Motion-aware Policy (TEMPO)", "role": "proposed_method", "latex": "\\mathbf{c}_{t}=\\bigl(\\bar{q}^{(1)}_{t},\\dots,\\bar{q}^{(K)}_{t}\\bigr),\\qquad\\bar{q}^{(k)}_{t}=\\frac{1}{L}\\sum_{\\tau\\in B_{k}}q_{\\tau}\\in\\mathbb{R}^{d_{q}},\\qquad q_{\\tau}:=0\\ \\text{for }\\tau<\\text{episode start}.", "mathml": "𝐜t=(qΒ―t(1),…,qΒ―t(K)),qΒ―t(k)=1Lβ€‹βˆ‘Ο„βˆˆBkqΟ„βˆˆβ„dq,qΟ„:=0​for ​τ<episode start.", "context": "Raw proprioceptive history is long, high-frequency, and largely redundant, so we compact it into a fixed-size summary 𝐜t (TEMPOACT). Let the proprioceptive command at time Ο„ be qΟ„βˆˆβ„dq. We consider the most recent H commands preceding time t, partition them into K contiguous temporal buckets B1,…,BK of equal length L (so that H=K​L), ordered from oldest to newest, and summarize each bucket by its mean:", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" } ], "equations_status": "selected", "core_kind": "formula", "equations_fetched_at": "2026-10-16", "equations_source_url": "", "equations_source_sha256": "", "exposures_reviewed_at": "", "exposures_source_sha256": "" }