PaperlayerView exact revision on arXiv

Source-linked AI summary

Robota-Embed: Unified VLA Embeddings for Robot In-Context Learning and Evolution

Yunze Liu, Jing Wang, Shenrui Wu, Wenkun He, Guanqi Zhan, Xianzheng Ma, Junyu Chen, Wenhang Ge, Haozhan Tang, Yuchao Gu, Zhuoyang Zhang, Zekai Wang, Dantong Niu, Zhen Dong, Yuke Zhu, Trevor Darrell, Linxi "Jim" Fan, Song Han, Han Cai

arXiv:2610.19495v1cs.ROcs.CV

TL;DR

Robots need retrieved experience that matches task intent and execution stage to supply an appropriate continuation. Robota-Embed jointly encodes vision, language, and action at current-state and recorded-continuation endpoints. Its retrieval space supports in-context policy conditioning and memory evolution, with the evolved harness outperforming direct policy execution across five ManiSkill tasks.

  • Problem

    Robot experience retrieval must match task intent and execution stage because similar objects and instructions can call for different continuations.

    Source: §1 · p. 2

  • Method

    Robota-Embed jointly encodes vision, language, and native action in current-state queries and recorded continuations, using predictive retrieval training to align the endpoints.

    Source: §1 · p. 2, §3.2 · p. 3

  • Results

    Across five ManiSkill tasks, Embed2Evolve achieved 68% success versus 24% with GPT-6 Astra Direct using the selected context harness.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The shared retrieval space supports Embed2ICL policy conditioning and Embed2Evolve memory evolution around a frozen policy.

    Source: §1 · p. 2, §7 · p. 10

    Retraining and masking comparisons do not establish a universal modality ranking: masking reflects the trained model's input use, not attainable separately trained unimodal performance.

    Source: §6 · p. 10

Abstract

from arXiv · show

Retrieving useful context for a robot requires matching its task, execution stage, and the continuation that should follow. We introduce Robota-Embed, to our knowledge the first vision-language-action (VLA) embedding model to jointly encode all three modalities in both current-state queries and recorded future continuations. Predictive retrieval training aligns these endpoints, enabling historical continuations to serve as context at inference time. We construct a multi-source corpus of 1,695,684 cleaned training anchors and introduce Robota-Embed-Bench, with 4,000 queries and a shared 8,000-entry gallery. Robota-Embed achieves 47.33% clip-level and 81.77% episode-level R@1, exceeding action-free Gemini Embedding 2 by 22.09 and 33.47 percentage points. Controlled retrieval studies show that removing native action degrades retrieval more than removing vision. We study two complementary applications. Embed2ICL uses retrieved experience for training-free and training-based policy conditioning, increasing success from 47.0% to 56.3% in DROID-Sim and from 86.4% to 92.8% on the LIBERO difficult subset, while improving task progress on real robots. Embed2Evolve uses execution feedback to grow task-local memory and refine context use around a frozen GPT-6 Astra policy. Across five ManiSkill tasks, it achieves 68% success, compared with 24% for GPT-6 Astra Direct and 33% for a harness independently evolved with Omni-Embed-Nemotron-3B under the same protocol. Relative to the latter, it uses approximately 79.5% fewer policy tokens per test rollout, counting both successes and failures. These results show that a unified, action-aware retrieval space can support robot in-context learning and task-local context evolution.

1. Introduction

Robot experience retrieval must match task intent and execution stage to find a suitable continuation, not merely a similar scene. Robota-Embed jointly represents vision, language, and native action for retrieval, supporting demonstration-conditioned policies and memory evolution.

1. Introduction

  • A useful continuation must match task intent and execution stage: similar objects and instructions can require different next actions.The query describes the present, while a candidate records what followed an earlier interaction.Source: §1 · p. 2
  • Robota-Embed jointly encodes vision, language, and native action in both current-state queries and recorded continuations, with source-specific projectors for heterogeneous action formats.Its retrieval training combines predictive alignment, multimodal grounding and discrimination, and representation regularization.Source: §1 · p. 2
  • 47.33% clip-level and 81.77% episode-level R@1 across four sources demonstrate Robota-Embed's retrieval performance.Robota-Embed-Bench contains 4,000 queries and 8,000 gallery candidates spanning human, simulated, and real-robot experience.Source: §1 · p. 2
  • With backbone, data, and training budget matched on DROID, adding native action raises clip-level R@1 from 59.8% to 78.7%.Retraining and query-masking studies examine what information the representation learns and uses.Source: §1 · p. 2
  • Embed2ICL's two policy interfaces improve simulation task success and real-robot task progress; random-reference controls show that reference selection matters.DreamZero uses retrieved videos to initialize its visual cache without policy updates; Fast-WAM is fine-tuned to condition on precomputed Context Tokens.Source: §1 · p. 2
  • 68% success across five ManiSkill tasks, versus 24% with GPT-6 Astra Direct, came from the selected Embed2Evolve context harness.It used 43.7% fewer policy tokens per test rollout than Direct; against an independently evolved Omni-Embed-Nemotron-3B harness with 33% success, it used approximately 79.5% fewer tokens.Source: §1 · p. 2

2. Related Work

Prior work develops multimodal embeddings, physical-experience retrieval, and robot in-context learning or memory. Robota-Embed differs by using vision, language, and native action in both retrieval endpoints.

Multimodal and action-aware embeddings

  • Robota-Embed jointly encodes vision, language, and native action at both the current-state query and recorded-future-clip endpoints.Prior work includes image/video and multi-sensory embeddings, action-aligned representations, and VLA retrieval using action-token candidates or policy features.Source: §2 · p. 3

Physical-experience retrieval

  • Robota-Embed includes native action in both query and candidate keys, so motion informs similarity rather than only the retrieved content.Prior approaches variously retrieve proprioceptive sequences, actions selected through action-free keys, action–effect evidence, or policy features.Source: §2 · p. 3

Robot In-Context Learning and Evolution

  • Embed2ICL retrieves experience for WAM conditioning, while Embed2Evolve grows task-local memory and refines retrieval and context settings around a frozen policy.Unlike episode-reset memory, Embed2Evolve evaluates selected settings on fresh task initializations without further changes.Source: §2 · p. 3

3. Robota-Embed

Robota-Embed retrieves recorded continuations by matching a current robot state to future experience. Its unified VLA representation is trained with retrieval objectives and evaluated using a multi-source corpus and benchmark.

3.1. Problem Formulation

  • A query combines the instruction, interaction-region image, and synchronized native action; its target pairs the same instruction with a four-second future video clip.The future window follows the accepted anchor.Source: §3.1 · p. 3

3.2. Action-Native Predictive VLA Embeddings

  • Robota-Embed uses shared multimodal components, endpoint-specific adapters, and source-specific action projectors to encode each endpoint as a normalized vector.Training combines predictive alignment, multimodal grounding and discrimination, and representation regularization.Source: §3.2 · pp. 3–4, Figure 2 · p. 4

3.3. Training Data and Retrieval Benchmark

  • The corpus contains 1,695,684 filtered current–future pairs from four sources, and the benchmark tests 4,000 queries against a shared gallery of 8,000 clips.Clip-level retrieval accepts only the paired future clip, whereas episode-level retrieval accepts any clip from the same source episode.Source: §3.3 · pp. 4–5, Figure 3 · p. 4

4. Embedding Evaluation

Robota-Embed leads the four-source retrieval benchmark at both clip and episode levels. Controlled studies examine benchmark validity, representation adaptation, modality use, and the distinction between episode matching and continuation localization.

4.1. Unified VLA Retrieval on Robota-Embed-Bench

  • 47.33% clip-level and 81.77% episode-level R@1: Robota-Embed exceeds Gemini Embedding 2 by 22.09 and Omni-Embed-Nemotron-3B by 33.18 percentage points.The comparison uses the shared 8,000-clip gallery.Source: §4.1 · p. 5, Table 1 · p. 5
  • Masking candidate text lowers clip-level R@1 from 47.33% to 12.78%, while same-instruction galleries still yield 47.62% R@1.The same-instruction result indicates retrieval distinguishes continuations beyond instruction matching.Source: §4.1 · p. 5
  • Expanding the gallery from 1M to 15M entries raises query-encoding-plus-retrieval latency from 160.36 to 195.17 ms on one NVIDIA A100.These timings exclude preprocessing, reference reranking, and downstream policy execution.Source: §4.1 · p. 5

4.2. What Makes Predictive Retrieval Work?

  • On a matched DROID split, adapting the backbone and head reaches 78.7% R@1 versus 13.8% for a learned head over frozen features.Predictive alignment reaches 75.6% R@1; regularization and multimodal grounding/discrimination raise it to 77.0% and 78.7%.Source: §4.2 · p. 6
  • Removing action lowers R@1 by 18.9 points versus 1.9 without vision in retraining; masking query action or vision lowers clip R@1 by 27.25 or 24.33 points.Retraining measures what remaining modalities can relearn, while masking measures reliance under an input shift.Source: §4.2 · p. 6

4.3. Benchmark Retrieval Examples

  • Figure 4 contrasts a paired-clip hit with a same-episode near-miss, showing that episode matching does not guarantee localization of the paired continuation.In the near-miss, the paired clip ranks seventh although a nearby clip from the same episode ranks first.Source: §4.3 · p. 6, Figure 4 · p. 6

5. From Embeddings to Robot Context

Embed2ICL conditions policies on retrieved demonstrations, while Embed2Evolve grows task-local memory and refines context use around a frozen policy. The reported simulation results show improved task success, and real-robot evaluations report task-progress gains.

5.1. Embed2ICL: Embedding-Guided Robot In-Context Learning

  • In DROID-Sim, Embed2ICL averages 56.3% success versus 47.0% without warm-up and 36.0% with random references.Both retrievers use the same reranking agent, and reranking improves both.Source: §5.1 · p. 6, Table 2 · p. 7
  • On the LIBERO difficult subset, retrieved Context Tokens raise average success from 86.4% to 92.8% across five tasks.The retrieved-context condition improves every task; same-checkpoint retrieved tokens outperform random-reference tokens by 4.8 percentage points.Source: §5.1 · p. 7, Table 3 · p. 8
  • Both Embed2ICL interfaces improve real-robot task progress, with Fast-WAM gaining more on Split Cup than on Pick Fruit.Scores reflect task-specific completion stages and are not directly comparable across tasks.Source: §5.1 · p. 7, Table 5 · p. 8, Figure 6 · p. 8

5.2. Embed2Evolve: Embedding-Guided Context Evolution

  • Embed2Evolve grows task-local memory from successful and failed attempts and refines retrieval, refresh, and retained-history settings while keeping the policy and embedding weights fixed.Development selects the memory and context recipe, which are then fixed for fresh test initializations.Source: §5.2 · pp. 8–9, Figure 7 · p. 9
  • 68% success across five ManiSkill tasks: Embed2Evolve exceeds GPT-6 Astra Direct at 24% and an independently evolved Omni-Embed-Nemotron-3B harness at 33%.It uses approximately 79.5% fewer policy tokens per test rollout than the Omni-Embed-Nemotron-3B harness.Source: §5.2 · p. 9, Table 6 · p. 9
  • Compared with Direct, Embed2Evolve uses 43.7% fewer tokens across test rollouts and reduces reference API cost by 41.8%, from $9.9641 to $5.7980.Observed execution-time savings are not an isolated method speedup and are affected by caller and scheduling differences.Source: §5.2 · p. 10

6. Discussion

Useful robot context depends on matching task, execution stage, and continuation, while retrieval quality and context use remain subject to important scope limits. The discussion distinguishes experience identification from continuation localization and outlines future extensions to embodiments and temporal windows.

6. Discussion

  • Episode recall identifies experience, while clip recall measures continuation localization; their gap can leave the execution phase unresolved.Same-instruction evaluation and candidate-text masking probe temporal discrimination and target-language use beyond episode identification.Source: §6 · p. 10
  • Language specifies the task, action supplies motion and phase, and vision localizes interaction; masking measures the trained model’s input use, not separately trained unimodal ceilings.The no-vision retraining/masking gap supports a joint VLA key, not a universal modality ranking.Source: §6 · p. 10
  • Retrieved context helps only with a compatible interface: random references hurt DreamZero, while retrieved demonstrations improve DreamZero and Fast-WAM outcomes.The comparisons do not isolate the causal contribution of reference selection, temporal content, or conditioning individually.Source: §6 · p. 10
  • Embed2Evolve can grow task-local memory while retaining less history and refreshing less often; fewer policy calls need not mean shorter robot motion.Token and reference-cost metrics cover all test rollouts, while decision/time comparisons use shared successes.Source: §6 · p. 10
  • Future work targets more embodiments with diverse kinematics and action spaces, plus variable-length continuations instead of a fixed four-second window.The stated goal is to retrieve only the experience duration needed for the current interaction.Source: §6 · p. 10

7. Conclusion

The conclusion presents Robota-Embed as a shared space for task intent, visual state, and native action, with action improving retrieval beyond language and vision. It connects that representation to Embed2ICL and Embed2Evolve, including reported success and efficiency gains around a frozen policy.

7. Conclusion

  • Robota-Embed jointly encodes task intent, visual state, and native action for physical-experience retrieval.The conclusion states that action improves retrieval beyond language and vision.Source: §7 · p. 10
  • The shared space supports Embed2ICL conditioning and Embed2Evolve memory evolution.Under matched settings, the reported Embed2Evolve gains concern success and efficiency around a frozen policy.Source: §7 · p. 10

A.1. Architecture and Optimization

Robota-Embed builds endpoint embeddings from synchronized current-state queries and strictly future continuations, using shared multimodal processing with source-specific action handling. Its training objective, curated data, and benchmark design target retrieval of task-relevant experience and its temporal continuation.

A.1. Architecture and Optimization

  • Shared Transformer and projection process query and target endpoints, while separate residual adapters specialize current-state and future-clip readouts.Each endpoint pools multimodally fused hidden states at native action-token positions; source-specific projectors preserve native action formats.Source: §A.1 · p. 13

A.2. Robota-Embed Objective

  • The training objective combines predictive alignment, multimodal grounding and discrimination, and representation regularization.The final retrieval checkpoint follows 26,275 optimizer steps with equal sampling across four sources.Source: §A.2 · p. 13, §A.1 · p. 13, Table 8 · p. 14
  • Positive pairs link a current state to its paired future clip, while negatives require explicit semantic conflict and validity filtering.Same-episode near-duplicate windows and cross-episode action negatives with matching future gripper phase are excluded.Source: §A.2 · p. 14
  • Action remains part of fused retrieval alignment, but the objective omits standalone action-only alignment because one current action does not uniquely determine a future trajectory.Modality swaps replace one input while keeping the paired target fixed, with constrained donor conditions for text, image, and action.Source: §A.2 · pp. 14–15

A.3. Training Data and Curation

  • The training corpus contains 1,695,684 filtered anchors from EgoDex, DROID, LIBERO, and RoboMIND–Franka-dual, partitioned by episode, session, and content identity.Accepted anchors pair with strictly future four-second action sequences and 16 video frames; crop verification checks object visibility and match to the instruction.Source: §A.3 · pp. 15–16

B.1. Robota-Embed-Bench: Composition and Evaluation

  • Robota-Embed-Bench has 4,000 queries and a shared 8,000-candidate gallery, including paired future clips and nearby same-episode clips.Those nearby clips often share instruction, objects, and visual content, requiring temporal localization; evaluation uses cosine similarity without reranking.Source: §B.1 · p. 16
  • External baselines use their supported input interfaces, while the controlled DROID study matches training data, backbone architecture, and optimization budget to isolate modality choices.Robota-Embed uses instruction, a SAM3 object zoom, and synchronized action for queries, and instruction, future video, and future action for gallery entries.Source: §B.1 · p. 16

B.2. Per-Domain Results

Per-domain results show a larger gap between episode and clip retrieval on EgoDex than on DROID, while external baselines also vary across sources. The section reports both retrieval levels and emphasizes that identifying an episode may not locate its paired continuation.

B.2. Per-Domain Results

  • Recall and MRR are percentages, and lower mean ranks are better; source averages are equally weighted.Episode-level metrics first average within episodes and then across 800 episodes per source.Source: §B.2 · p. 17
  • 87.06% episode-level R@1 versus 36.70% clip-level R@1 on EgoDex, compared with 99.88% versus 74.10% on DROID.The larger EgoDex gap indicates that episode identification often leaves the paired window unresolved.Source: §B.2 · p. 17, Table 10 · p. 17
  • VLM2Vec-V2 and WeMM-Embedding-2B both perform best on DROID and EgoDex, with markedly lower recall on LIBERO and RoboMIND-FD.Table 11 reports action-free VL baseline recall and MRR by source; the supplied text does not give the corresponding cell values.Source: §B.2 · p. 17, Table 11 · p. 17

B.3. Candidate Language and Gallery Controls

The controls test whether retrieval uses candidate language and can localize continuations among same-instruction clips, then examine how retrieval training, trainable scope, objectives, and visual inputs affect performance.

B.3. Candidate Language and Gallery Controls

  • 47.62% clip-level R@1 on same-instruction galleries shows retrieval distinguishes continuations sharing a task instruction; candidate-text masking instead lowers R@1 from 47.33% to 12.78%.The same-instruction results are 38.60% on LIBERO and 40.10% on RoboMIND-FD, above random rates of 1.90% and 0.24%.Source: §B.3 · p. 18, Table 13 · p. 18, Table 14 · p. 18

B.4.1. Retrieval Adaptation

  • 47.33% clip-level and 81.77% episode-level R@1 follow retrieval training, compared with 0.65% and 1.17% from frozen Cosmos3-Edge features.The four-source benchmark averages the shared Robota-Embed-Bench gallery.Source: §B.4.1 · p. 19, Table 15 · p. 19

B.4.2. Trainable Scope

  • 78.7% R@1 with a trained backbone and head exceeds 13.8% from training a retrieval head over frozen features on the matched DROID split.Both variants use a learned retrieval head; the higher-scoring variant also updates the backbone.Source: §B.4.2 · p. 19, Table 16 · p. 19

B.4.3. Objective Groups

  • 77.0% R@1 with regularization and 78.7% with added multimodal grounding and discrimination improve on 75.6% from predictive alignment alone.R@10 is 100.0% in all three runs.Source: §B.4.3 · p. 19, Table 17 · p. 19

B.4.4. Visual Input

  • 78.7% R@1 and 88.00% MRR with SAM3 object zoom exceed 77.9% and 87.40% with full frames and 76.8% and 86.69% without vision.The object-zoom R@1 gain over full frames is 0.8 points.Source: §B.4.4 · p. 20, Table 18 · p. 20

B.4.5. Training-Time Modalities

Controlled modality tests compare retraining without inputs against masking inputs in a fixed model. The remaining results describe retrieval scaling, qualitative cross-embodiment examples, and Embed2ICL interfaces and evaluations.

B.4.5. Training-Time Modalities

  • In controlled DROID retraining, removing language has the largest effect; removing action lowers R@1 from 78.7% to 59.8% (18.9 percentage points), versus a 1.9-point drop without vision.Source: §B.4.5 · p. 20

B.5. Controlled DROID Query Masking

  • Masking action, image, or text in the fixed DROID model lowers clip-level R@1 from 78.7% to 50.7%, 60.9%, or 1.9%, respectively.Retraining allows adaptation to remaining inputs, whereas masking measures fixed-model reliance under input shift.Source: §B.5 · p. 21, Table 21 · p. 21

B.6. Index-Size Scaling and Retrieval Latency

  • Expanding the precomputed gallery from one million to 15 million entries adds 34.81 ms to Top-10 search latency; query encoding takes 157.70 ms at every scale.Measurements use one NVIDIA A100 and exclude preprocessing, reranking, policy inference, and robot execution.Source: §B.6 · p. 21, Table 22 · p. 21

B.7. Qualitative Cross-Embodiment Analysis

  • Three selected diagnostic cases show top-ranked clips sharing placement or removal roles with queries across human, robot, and simulated experience.The examples are qualitative evidence from a separate checkpoint, not aggregate evidence of accuracy, robustness, or source invariance.Source: §B.7 · p. 21, Figure 8 · p. 22

C.1. DreamZero-DROID Reference Warm-Up

  • DreamZero-DROID receives a retrieved, reranked four-second video to warm its visual cache, with the policy frozen during execution.References use three synchronized streams at 15 Hz and remain fixed across trials for each task; Figure 10 shows task-specific examples preserving placement relations.Source: §C.1 · p. 22, §C.2 · p. 22, Figure 9 · p. 23

C.3. Training-Free Embed2ICL on the YAM Robot

  • On the YAM robot, retrieved-reference conditioning raises mean progress from 21.67 to 40.00 and success from 0.0% to 16.7% across 12 rollouts per condition.The adapted policy is shared across conditions, and the retrieved reference depicts placement into a bowl.Source: §C.3 · pp. 23–24

C.4. Fast-WAM Integration Details

  • Fast-WAM retrieves precomputed Context Tokens and appends their projected sequence to native text and proprioceptive conditioning for cross-attention.Offline caching stores retrieval vectors, tokens, and masks; the online query selects a top-1 reference. Policy adaptation updates approximately 6.02B parameters while the retrieval encoder remains offline.Source: Figure 11 · p. 24, §C.4 · pp. 24–25

C.5. LIBERO Supplementary Controls

  • 92.8% LIBERO difficult-subset success with retrieved tokens exceeds 88.0% with random-reference tokens and 86.4% after matched no-ICL continuation.The retrieved-versus-random comparison holds policy weights fixed; continuation for 21,700 updates does not reproduce the retrieved-context result.Source: §C.5 · p. 25

C.6. Training-Based Integration Ablation

The section compares two ways to condition Fast-WAM and then describes how Embed2Evolve develops and tests task-specific context harnesses. It reports integration outcomes and clarifies the scope of the embedding-model comparison.

C.6. Training-Based Integration Ablation

  • 92.8% average LIBERO success for prefilling exceeds 88.0% for the frozen-backbone side branch and 86.4% for the official release.The side branch performs better on L6 (43 versus 42 successes), while prefilling performs better on four tasks; the side branch uses fewer trainable parameters.Source: §C.6 · p. 26, Table 23 · p. 26
  • On real robots, prefilling reaches 79.17% progress on Pick Fruit and 87.50% on Split Cup, versus 77.08% and 75.00% for the side branch.Both settings improve mean normalized task progress over no ICL.Source: §C.6 · p. 27, Table 24 · p. 27
  • Retrieved tokens outperform random tokens in both integrations: 92.8% versus 88.0% for prefilling and 88.0% versus 85.6% for the side branch.For prefilling, both reference conditions use the same policy checkpoint.Source: §C.6 · p. 27, Table 25 · p. 27

D.1. Policy and Memory Interfaces

  • Embed2Evolve keeps the policy and embedding models fixed while task-local memory grows and retrieval, refresh, and retained-history settings are refined.The matched embedding-model comparison independently develops each system using the same policy, prompt, initial memory, seeds, candidate rules, and stopping criteria.Source: D. Embed2Evolve: Harnesses,… · p. 27, §D.1 · p. 27

D.2. Development and Selected Harnesses

  • Configuration selection prioritizes success count, then fewer control ticks, shorter execution time, and the earlier round; search also varies retrieval count, refresh cadence, history, and scoring bonuses.The selected configuration and memory bank are frozen for formal testing, whose outcomes do not enter selection.Source: §D.2 · pp. 29–30

D.3. Development Trajectories

  • Robota-Embed’s selected configurations emerge from development runs of four rounds for StackCube, eight for TwoRobotStackCube, and the eight-round cap for three other tasks.The 36 recorded configurations and 360 trials include historical first rounds ineligible for selection; selected harnesses need not come from the final round.Source: §D.2 · p. 30, Table 28 · p. 30, §D.3 · p. 31

D.4. Embedding-Model Comparison

  • The embedding-model replacement changes both retrieval inputs and retrieval training, so the independently evolved systems are not an action-only ablation or an encoder swap within one fixed harness.Fewer high-level decisions can coexist with more physical control ticks because the measures capture different aspects of execution.Source: Table 29 · p. 31, §D.4 · p. 32

D.5. Fresh-Test Protocol and Efficiency

Embed2Evolve’s fresh-test efficiency comparisons use five tasks and 400 test episodes, with success and token means covering all trials but decision and time comparisons restricted to shared successes. The related-protocol comparison distinguishes retrieval inputs, action roles, policy coupling, and experience lifetime.

D.5. Fresh-Test Protocol and Efficiency

  • The fresh-test protocol uses 20 shared test seeds per task across four conditions, yielding 400 episodes; success and token means include all trials, while decision and time comparisons use shared successes.Budget exhaustion and policy failures count as failures; audited infrastructure interruptions are excluded.Source: §D.5 · p. 32, Table 30 · p. 32
  • 1.90× is the mean task-wise Direct-to-Evolve execution-time ratio, while the reported 41.4% reduction averages task-wise relative savings on shared-success pairs.Execution time includes API waits; uncontrolled caller handling, concurrency, CPU placement, and API load confound attribution, so the ratios are observed execution rather than isolated method-induced speedup.Source: §D.5 · p. 32, Table 31 · p. 33

D.6. Token Usage and Reference Cost

  • 43.7% fewer total tokens per rollout (920,979.76 to 518,216.15) and 41.8% lower reference cost ($9.9641 to $5.7980) are reported for Embed2Evolve versus Direct.Means weight the five task means equally and include all test rollouts, including failures.Source: Table 32 · p. 33, §D.6 · p. 33
  • Total token use decreases on every task, with the largest reduction, 57.5%, on Pull Cube with Tool; input-token reductions account for most savings.Stack Cube output usage increases, but its input-token reduction still lowers total consumption.Source: §D.6 · p. 33

E. Related Retrieval Protocols

  • The related-protocol comparison distinguishes whether action enters the retrieval key, retrieved record, or policy output, alongside policy coupling and retrieved-experience lifetime.For Zeva, both phase keys encode recurrent interaction history including executed actions and observed effects; Embed2ICL and Embed2Evolve share Robota-Embed’s retrieval space.Source: E. Related Retrieval Protocols · p. 34, Table 33 · p. 34

Related papers

Loading 2610.19495v1…