Being-H0.7: A Latent World-Action Model from Egocentric Videos
1. Paper overview
In one sentence: Being-H0.7 uses future-informed latent alignment to train a direct robot policy, gaining strong whole-system performance while leaving the contribution of each training ingredient unisolated. E02E03E04E07E12E13E14E15
| At a glance | What to know |
|---|---|
| Research problem | Author claim Sparse action supervision can encourage shortcuts in direct vision-language-action policies. The authors argue that reconstructing future pixels spends capacity on appearance irrelevant to control and increases cost. They seek predictive, action-useful representations that retain direct policy deployment. E02 |
| Core mechanism | |
| A key reported result | Real-robot ability suites: Dynamic 70.0; Physical 66.9; Motion 67.5; Long Horizon 66.7; Generalization 67.5 Suite success rate (%). Twelve tasks across PND Adam-U, Unitree G1, and Franka FR3; 20 blind trials/task/method; suites average all tasks with each tag. Best competing bars: Fast-WAM 58.3/57.5 for Dynamic/Motion; Being-H0.5 59.4/60.6 for Physical/Long Horizon; π0.5 58.3 for Generalization. Leads all displayed suites. Shared tasks make suite scores dependent; no per-task success counts or confidence intervals are supplied. E10E11E12 |
| Reading caution |
Core contributions
Figure 3. Two training branches share context while future information reaches the prior through latent alignment. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the gray shared-context tokens. The prior appends latent queries and noised actions; the posterior appends future embeddings and its own noised actions. Follow the vertical paths to the matched hidden states. The double-ended alignment arrow denotes a training loss, not direct attention between branches. On the right, the white cross-branch blocks agree with Section 3.3's prohibition on prior/posterior attention, while both branches retain access to context. Each action path has a flow-matching loss, and both latent paths receive regularization. Section 3.2 supplies the input detail absent from this drawing: future embeddings come from a frozen ViT and Perceiver resampler. E03E04E05E15
What it supports. The deployable policy learns from a privileged training counterpart without needing that counterpart's future observations during control. Sharing context and matching token positions provide an efficient implementation of this idea. The diagram supports an integrated policy architecture, not a separate video-rollout planner feeding actions to another controller.
Where the evidence stops. The alignment arrow must not be read as test-time access to future frames. Only the future visual backbone is explicitly frozen; the posterior hidden representation is jointly optimized. The supplied paper does not ablate branch alignment or regularization.
2. Motivation
2.1 The problem and the proposed response
Sparse action supervision can encourage shortcuts in direct vision-language-action policies. The authors argue that reconstructing future pixels spends capacity on appearance irrelevant to control and increases cost. They seek predictive, action-useful representations that retain direct policy deployment. E02
2.2 What this reading follows
A robot catching a moving ball needs useful anticipation before its next action, but it need not render a detailed movie. Being-H0.7 puts that idea into a policy with a compact set of reasoning queries. During training, a second branch sees future observations and aligns its hidden states with those of the deployable branch. At control time, only current context and the learned queries remain. The following visuals separate this information flow from three different kinds of evidence: benchmark performance, qualitative future visualization, and deployment cost. The paper reports encouraging results, but provides no component ablation that attributes those gains specifically to future alignment. E02E03E04E07E12E13E14E15
3. Research context
We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.
| Catalog dimension | Recorded classification |
|---|---|
| Major category | Not assigned |
| Architecture | Not assigned |
| Prediction paradigm | Not assigned |
| Quadrant | Not assigned |
This table preserves the labels recorded at reading time. The current major category is WAMs. View the current classification.
3.1 Evidence-based assessment
Insufficient evidence to decide
The catalog is unassigned. Architecture evidence supports an integrated MoT policy with Understanding and Action Experts and shared prior/posterior computation. Its mechanism is training-time future-latent alignment, with only the prior used for control. It does not emit joint future-video/action rollouts or use a separate inference-time inverse-dynamics model. This supports a latent predictive-policy interpretation; no recorded quadrant exists to affirm or contradict. E03E04E05
This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.
4. Problem formulation
4.1 Inputs and outputs
| Inputs | Outputs |
|---|---|
|
|
4.2 Equations and their role
5. Method in detail
5.1 Use future evidence to train a deployable reasoning interface
Begin with the information available before an action: instruction, observation history, and robot state. The prior places learnable queries after that context and before the action chunk. These positions participate in transformer propagation, giving action generation an intermediate representation to use. During training, the posterior keeps the same context and action pathway but replaces the queries with embeddings from subsequent observations. A frozen visual backbone and Perceiver produce embeddings with the same shape as the queries. The branches therefore offer matched locations at which to compare current-context reasoning with future-informed reasoning. The deployed system removes the posterior; it does not retrieve actual future observations or run the diagnostic video generator before choosing actions. E02E03E04E13
5.2 Understand what alignment encourages—and what it cannot guarantee
Equation (3) penalizes differences between the branches' corresponding latent hidden states across aligned layers. Both branches also learn action generation through flow matching, so the future-informed representation is tied to the demonstrated action target. My interpretation is that this encourages the interface to retain future-related information that is useful for action and inferable from current context. That interpretation is a learning rationale, not a measured causal result. A trivial way to lower alignment error would be to shrink or homogenize representations. The paper therefore adds a norm threshold and a spectral-diversity penalty during pretraining, then removes those regularizers during task-specific post-training. Because there is no component ablation or measured collapse diagnostic, the supplied evidence does not show how much each safeguard contributes. E03E04E05E06E15
5.3 Follow predictions through buffering to physical execution
A generated action chunk is only one part of a robot controller. Here the client consumes actions at the platform's policy frequency while UAC requests another chunk using fresh observations. Its committed prefix stays fixed, and newly returned actions update only a future suffix after an estimated delay. Unitree G1 adds an AMO whole-body backend, making the distinction between upper-body policy targets and stable execution especially concrete. My reading is that the success bars test this assembled system, whereas the cost bars characterize its deployment scheduling. Neither supplies a clean estimate of the causal benefit of latent alignment. A useful reproduction must therefore hold both model training and observation-to-execution timing under control rather than treating a low latency number as proof of better prediction. E10E11E12E14E15
5.4 Training and inference
During training
Both branches predict flow velocities from noised actions. Pretraining adds norm preservation and spectral-diversity penalties: these discourage vanishing states and minimize negative entropy of a normalized Gram-matrix spectrum to discourage directional collapse. These effects are not verified by component ablations. E05E15
Built on Being-H0.5, the model uses InternVL3.5/Qwen3 experts and V-JEPA2.1 encoders; the context encoder stays trainable. Context/future images are 224×224/256×256; H=4, T=20, K=16, and the last L=9 layers are aligned. Loss weights are 10⁻³ for alignment and 10⁻⁴ each for norm/rank regularization. Task-specific post-training drops regularization, retaining action/alignment losses at approximately 128 chunks per effective global batch. E03E06
During inference
Discard the posterior and generate actions using current instruction, observations, state, and queries. Universal Async Chunking (UAC) executes a committed action prefix while a parallel thread requests another chunk from fresh observations; only the future suffix is updated after the estimated delay. E02E03E14
Execution uses embodiment-specific interfaces. Unitree G1 additionally relies on pretrained AMO for balance-aware whole-body control; policy prediction is distinct from the complete low-level execution stack. E10
5.5 Implementation flow
- Insert reasoning positions
The sequence places K learnable queries Q between instruction/observation/state context and the action chunk. Transformer propagation organizes these states to guide actions; simply adding queries does not guarantee predictive content. E03
- Construct the privileged branch
Replace posterior queries with same-shaped future embeddings. A frozen pretrained ViT encodes future observations; a Perceiver resampler aggregates them into K embeddings. Both branches share the backbone and action-generation pathway. E03
- Align without attention leakage
Pack both branches into one MoT sequence. The Action Expert processes action/state vectors; the larger Understanding Expert processes other signals. Both branches access shared context but cannot attend to each other. Corresponding positions share positional IDs; matched latent hidden states receive squared-distance alignment. E04
6. Experiments & results
Being-H0.7 trains a robot policy to anticipate useful future information inside latent queries. A future-aware training branch aligns its hidden states with a deployable branch that sees only current context, avoiding visual rollout during control. Strong benchmark and real-robot results support the complete system, while missing component ablations leave the causal contribution of future alignment unresolved (E02–E09, E12, E15).
The complete supplied v1 contains an architecture figure, quantitative comparisons, a qualitative latent-video diagnostic, and a deployment-cost figure, but no component ablation or quantitative latent-prediction diagnostic. Figure 7 is included as a qualitative diagnostic, not relabeled as measured ablation evidence. The author-only appendix adds no experiments. Separate supplemental availability remains unverified. E13E15
6.1 Read the original evidence
Table 1. Strong cross-benchmark results include both leading scores and clear remaining gaps. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read vertically within a benchmark column before comparing model families. The final row is the 3B Being-H0.7 model. Preserve the protocol definitions above the table: CALVIN means ABCD→D, CALVIN* means ABC→D, and LIBERO-plus* uses target-benchmark fine-tuning. Most performance columns are success percentages; CALVIN reports consecutive tasks out of five, and RoboTwin's paired numbers refer to Easy/Hard settings. A dash is a missing reported result, not zero performance. The model-size column and incomplete coverage make this a broad literature comparison rather than a controlled architectural experiment. The original short caption is retained because it defines the starred columns. E06E07E08E09E15
What it supports. Being-H0.7 reports 99.2% on LIBERO and 82.1% on zero-shot LIBERO-plus. The same table shows limits to a universal-lead claim: RoboCasa's 62.1% trails Cosmos-Policy's 67.1%, GR1's 49.2% trails ABot-M0's 58.3%, and RoboTwin's 90.2/89.6 trails the two strongest listed world-model rows.
Where the evidence stops. Do not average CALVIN task counts with success percentages or combine zero-shot and fine-tuned LIBERO-plus. Model sizes differ, matched pretraining is not established, many entries are missing, and the table reports no uncertainty estimates.
Figure 6. Being-H0.7 leads five overlapping real-robot ability suites under the reported blind protocol. Original paper, p. 12 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Use the legend below the bars: dark blue denotes Being-H0.7, light blue Being-H0.5, pale yellow pi0.5, and salmon Fast-WAM. Read the printed values as success percentages; the crop contains the complete legend and every suite label. The experimental unit is a task rollout, with twenty blind trials per task and method. Table 3 and Section 4.3.1 explain why a suite is not an independent task set: each includes all tasks carrying that primary or overlap tag. Thus a conveyor task can contribute to several groups. Compare models within a group, then consult the tagging and evaluation pages before interpreting consistency across groups. E10E11E12E15
What it supports. Being-H0.7 reaches 70.0%, 66.9%, 67.5%, 66.7%, and 67.5% from left to right. In Dynamic Scene, the nearest competing bar is Fast-WAM at 58.3%, an 11.7 percentage-point difference between displayed values. These are physical-execution results, providing a different kind of evidence from generated future images.
Where the evidence stops. Five leading bars do not constitute five independent replications: suite membership overlaps. No error bars or per-task outcome counts are supplied, and the comparisons do not isolate the latent-alignment mechanism from other model differences.
Figure 8. UAC changes measured deployment latency while each Being-H model retains its memory footprint. Original paper, p. 13 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the two panels independently because the model order differs. The left panel uses milliseconds per step; lower values indicate lower reported system-level latency. The right panel uses GPU memory in gigabytes. Locate Being-H0.7-UAC and Being-H0.7 by their labels rather than their positions, then compare their equal memory bars with their different latency bars. Section 4.3.3 describes UAC as asynchronous chunk buffering: actions are consumed from a locked prefix while an inference thread requests new predictions and replaces only a future suffix. This scheduling mechanism explains why the figure concerns the deployment stack, not just transformer evaluation time. E10E14E16
What it supports. Being-H0.7-UAC is reported at 3.61 ms/step, compared with 28.81 ms/step for Being-H0.7; both use 6.8 GB. Being-H0.5-UAC is slightly lower at 3.45 ms/step and uses 5.6 GB. The result supports efficient deployment with buffering, while showing that H0.7 adds memory relative to its predecessor.
Where the evidence stops. The source does not specify the GPU model or exact latency-accounting boundary. These values cannot be converted into isolated forward-pass speedups or new robot control frequencies; Table 2 reports the actual policy frequencies separately.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| LIBERO manipulation All four suites used for training; 500 evaluation trials per suite. | 99.2% Mean success rate | Being-H0.5: 98.9%; π0.5: 96.9%. Highest listed score, with a 0.3 percentage-point margin over Being-H0.5; no uncertainty is reported. E07 |
| LIBERO-plus robustness Standard-LIBERO training versus separate LIBERO-plus fine-tuning configuration. | 82.1% zero-shot; 84.8% fine-tuned Success rate | Being-H0.5: 78.5%/83.1%; MINT-4B: 80.1%/84.1%. The fine-tuned result uses target-domain training and must remain separate from zero-shot robustness. E07 |
| RoboCasa few-shot manipulation 24 tasks; 50 human demonstrations/task; 50 trials/task in held-out scenes. | 62.1% Success rate | Being-H0.5: 53.5%; Cosmos-Policy: 67.1%. Improves the predecessor but trails the strongest listed comparator. E08 |
| GR1 dexterous manipulation 24 bimanual tasks; 1,000 demonstrations/task; 50 evaluation trials/task. | 49.2% Mean success rate | gr00t-N1.6: 47.6%; ABot-M0: 58.3%. Competitive but not best in this humanoid benchmark. E08 |
| RoboTwin 2.0 manipulation 2,500 clean and 25,000 randomized demonstrations; Easy/Hard evaluation, 100 rollouts/task. | 90.2% / 89.6% Success rate, Easy/Hard | LingBot-VA: 92.9%/91.6%; Fast-WAM: 91.9%/91.8%. The within-model gap is 0.6 percentage points; both listed world-model baselines score higher. E09 |
| CALVIN instruction chains ABCD→D and ABC→D; 1,000 five-instruction sequences. | 4.67 / 4.48 Mean consecutive tasks completed, maximum 5 | Being-H0.5: 4.63/4.48. Improves seen-environment performance slightly and ties the predecessor on unseen D; these are task counts. E09 |
| Real-robot ability suites Twelve tasks across PND Adam-U, Unitree G1, and Franka FR3; 20 blind trials/task/method; suites average all tasks with each tag. | Dynamic 70.0; Physical 66.9; Motion 67.5; Long Horizon 66.7; Generalization 67.5 Suite success rate (%) | Best competing bars: Fast-WAM 58.3/57.5 for Dynamic/Motion; Being-H0.5 59.4/60.6 for Physical/Long Horizon; π0.5 58.3 for Generalization. Leads all displayed suites. Shared tasks make suite scores dependent; no per-task success counts or confidence intervals are supplied. E10E11E12 |
| Deployment inference cost System-level real-world inference stack; UAC and non-UAC configurations. | Being-H0.7-UAC: 3.61 ms/step, 6.8 GB; Being-H0.7: 28.81 ms/step, 6.8 GB Reported average latency (ms/step); GPU memory (GB) | Being-H0.5-UAC: 3.45 ms/step, 5.6 GB; π0.5-RTC: 12.45 ms/step, 9.3 GB; Fast-WAM: 41.26 ms/step, 15.6 GB. Deployment-stack measurements, not an isolated forward-pass benchmark. GPU model and latency-accounting details are absent. E14 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Figure 7. A qualitative probe visualizes task futures from observations and prior hidden states. Original paper, p. 12 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each labeled strip separately. Visible examples involve a racket and ball, catching with a tube, pouring at a conveyor, package handling, shoe packing, and grocery sorting. Each strip places one green-bordered image before three orange-bordered images; the paper does not provide an explicit border-color legend. Section 4.3.3 explains the experiment's role: a separate video generation model receives both the current observation and the prior branch's latent hidden states and synthesizes future task states. These are illustrative outputs of a diagnostic decoder. They are not frames produced by Being-H0.7 as part of its normal control loop, nor a reported comparison with ground-truth future video. E13E15
What it supports. The examples are visually consistent with the authors' claim that prior hidden states can support future-oriented visualization. The conservative finding is that a video generator can be conditioned this way and produce the shown task scenes. The figure does not establish how much predictive information comes from the latents rather than the observation or generator.
Where the evidence stops. This diagnostic occupies the ablation section because no component ablation is supplied; it is not a controlled ablation itself. No predictive-accuracy metric, observation-only decoder control, or sufficient generator implementation details accompany these examples.
7. Analysis & limitations
7.1 What the evidence leaves open
No ablation isolates queries, alignment, regularization, or human-video pretraining. Table 1 compares different model sizes and does not establish matched pretraining; complete-system gains cannot establish which ingredient caused them. E06E07E15
Figure 7 conditions a separate video generator on observations and prior hidden states. Its qualitative futures do not measure predictive accuracy or show that latent information is necessary for control; no observation-only generator control is reported. E13
The blind real-world protocol improves comparison, but overlapping suites, absent per-task outcomes, and unpublished operational success thresholds restrict independent statistical and procedural checking. E11E12
7.2 Questions for discussion
- Would matched-capacity action-only queries retain the gains attributed to future alignment?
- Does the prior add predictive information beyond observations under a controlled video-decoder probe?
- How much dynamic-task success remains after matching observation age and action-buffer delays?
8. Reproducibility audit
8.1 Requirements and known gaps
Reconstruction needs the inherited checkpoints, visual encoders, Perceiver, mixed UniHand 2.0 trajectories, and downstream demonstrations. This PDF omits corpus size/mixing weights, optimizer schedule, training duration/hardware, future-frame sampling, norm threshold, and rank-projection dimension. Its appendix supplies authors only. E03E05E06E16
Deployment needs RGB/state synchronization, action mappings, UAC buffering, and the G1 AMO backend. Table 2 gives policy frequencies of 20/10/20 Hz for Adam-U/G1/FR3, distinct from G1's AMO 50 Hz body loop. Denoising-step count and timing instrumentation remain unspecified. E10E14E16
8.2 Proposed reproduction checks
The following checks are proposals motivated by the paper. They have not been run as part of this reading.
Check 1: A matched query-and-alignment ablation
Reader-proposed, not performed: start from the same inherited checkpoint and a fixed subset of manipulation trajectories. Compare an action-only policy with the same K=16 query positions against the complete dual-branch model, keeping data, action horizon, trainable parameter capacity, updates, and deployment settings matched. Include a dual-branch control whose future embeddings are shuffled across trajectories, plus a full-model variant without the two pretraining regularizers. Evaluate held-out task success and record query norms, Gram spectra, and alignment error over training; report repeated-seed uncertainty. Genuine future supervision should outperform shuffled supervision and action-only queries. Low alignment error without improved success or noncollapsed states would weaken the proposed mechanism. E03E04E05E06E15E16
Check 2: Separate latent quality from observation age and buffering
Reader-proposed, not performed: use the rolling-ball task and fixed scene initializations from the real-world protocol. Cross an aligned policy and a matched action-only-query control with UAC enabled/disabled. Use the same robot, cameras, server hardware, policy frequency, and action horizon; log image timestamps, server turnaround, committed-buffer length, action execution times, and binary outcomes. Add controlled response delays and report uncertainty across randomized blind trials. If the aligned policy retains a success advantage at matched observation-to-execution delay, that favors a representation benefit. If the advantage disappears after delay matching, scheduling or observation freshness is a plausible contributor. This also tests whether the reported ms/step metric predicts actual reaction quality. E03E10E11E12E14E15E16
8.3 Reading coverage
Visual audit: Original PDF pages 1–14 and 22 were rendered and actually viewed, covering the title/version, appendix author credits, all retained method/training/evaluation details, all figures and tables, and the source facts underlying proposed checks. Figures 1–8 and Tables 1–3 were visually read. Each of the five final original crops was separately viewed; the table retains its short protocol-defining caption, and charts retain their legends and labels. Figure 3's branch isolation and alignment markings were cross-checked against Sections 3.2–3.3 and Equation (3). No conflicting claim-relevant marker was found. Reference pages 15–21 were read in the complete text chunks, not visually inspected. No external video, code, or separate supplement was inspected.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 22. Appendix coverage: reviewed.
Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.
Text reading scope & known omissions
- Abstract (p. 1)
- 1 Introduction (pp. 2–3)
- 2 Related Work (pp. 3–4)
- 3.1 Latent Reasoning (pp. 4–5)
- 3.2 Joint Alignment with Future Information (pp. 5–6)
- 3.3 Efficient Dual-Branch Implementation (pp. 6–7)
- 4.1 Training details (pp. 7–8)
- 4.2 Simulation (pp. 8–9)
- 4.3 Real-world Experiments (pp. 9–14)
- 5 Conclusion (p. 14)
- References (pp. 15–21)
- Appendix: Author List (p. 22)
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- Identity/version: the title exactly matches the catalog. The title page credits BeingBeyond Team; the Appendix Author List names all nine catalog authors in the same order, marking the first four as equal contributors and Zongqing Lu as corresponding author. No institutional affiliation is explicitly stated. The artifact is arXiv:2605.00078v1 [cs.RO], 30 April 2026; its internal title-page date is 14 April 2026. These are distinct source dates, not evidence of a different edition. No other revision was supplied or compared (E01).
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout. This supplied extraction limitation was addressed by visually inspecting original pages and final crops.
- Separate supplemental material availability has not been fully verified.
- All seven supplied chunks were read, including references and the author-only appendix. Pages 1–14 and 22 were visually inspected; reference pages 15–21 were read as text only. No external videos, project website, code, or separate supplement was inspected, and no experiments were reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
E01PDF p. 1, title/byline, arXiv margin and internal date; p. 22, Appendix: Author List
Title matches exactly. Byline: BeingBeyond Team. Appendix: Hao Luo, Wanpeng Zhang, Yicheng Feng, Sipeng Zheng, Haiweng Xu, Chaoyi Xu, Ziheng Xi, Yuhui Fu, and Zongqing Lu. First four are equal contributors; Lu is corresponding author. PDF labels v1, 30 April 2026, and internally 14 April 2026.
Go to primary source ↓E02PDF pp. 2–3, Section 1, motivation and deployable-prior discussion
The authors motivate latent prediction by action-supervision sparsity and pixel-rollout overhead. Future observations and the posterior are used only during training.
Go to primary source ↓E03PDF p. 5, Sections 3.1–3.2, Equations (1)–(2); p. 4, Figure 2
Instruction x, observation context o_{−H:0}, state s, K×d queries Q, and action chunk a_{0:T} form the sequence. The posterior replaces queries with same-shaped future embeddings from a frozen ViT and Perceiver; branches share context, backbone, and action pathway.
Go to primary source ↓E04PDF p. 6, Figure 3, Equation (3), Section 3.3
Matched prior/posterior hidden states receive normalized squared Frobenius-distance alignment. A packed MoT pass shares context, isolates cross-branch attention, and assigns corresponding positions identical IDs; action/state and other signals use distinct experts.
Go to primary source ↓E05PDF p. 7, Section 3.3, Equations (4)–(8)
Both branches use flow matching on interpolated actions with Gaussian noise. Norm-threshold and normalized Gram-spectrum negative-entropy regularizers supplement alignment. Mixed human/robot trajectories use UniHand 2.0; text generation is not evaluated.
Go to primary source ↓E06PDF p. 7, Section 4.1; p. 8, paragraph immediately below Table 1
Specifies expert families, V-JEPA2.1, trainable context encoder, 224/256 image resolutions, H=4, T=20, K=16, L=9, and loss weights. Downstream training retains action/alignment objectives without regularizers, at approximately 128 chunks per effective batch.
Go to primary source ↓E07PDF p. 8, Table 1, LIBERO/LIBERO-plus/LIBERO-plus* columns and named model rows; Section 4.2.1 LIBERO and LIBERO-plus bullets
Being-H0.7 scores 99.2/82.1/84.8, Being-H0.5 98.9/78.5/83.1, π0.5 96.9/77.4/absent, and MINT-4B 98.7/80.1/84.1. LIBERO uses four training suites and 500 trials/suite; the starred robustness setting fine-tunes on LIBERO-plus.
Go to primary source ↓E08PDF p. 8, Table 1, RoboCasa-50 and GR1 columns; Section 4.2.1 RoboCasa and GR1 bullets
RoboCasa: 24 tasks, 50 demonstrations and trials/task, held-out scenes; Being-H0.7 62.1, Being-H0.5 53.5, Cosmos-Policy 67.1. GR1: 24 tasks, 1,000 demonstrations and 50 trials/task; Being-H0.7 49.2, gr00t-N1.6 47.6, ABot-M0 58.3.
Go to primary source ↓E09PDF p. 8, Table 1, CALVIN/CALVIN*/Robotwin2 columns; p. 9, Section 4.2.1 RoboTwin and CALVIN bullets, Section 4.2.2
RoboTwin clean/randomized training: 2,500/25,000 demonstrations and 100 rollouts/task. Scores: 90.2/89.6 versus LingBot-VA 92.9/91.6 and Fast-WAM 91.9/91.8. CALVIN: 1,000 five-task sequences; Being-H0.7 4.67/4.48 and Being-H0.5 4.63/4.48 for ABCD→D/ABC→D.
Go to primary source ↓E10PDF p. 9, Section 4.3 and Figure 4; p. 10, Table 2 and Section 4.3.1 AMO paragraph
PND Adam-U, Unitree G1, and Franka FR3 use Linkerbot O6 hands, total DoF 31/26/13 and policy frequencies 20/10/20 Hz. G1 uses pretrained AMO at 50 Hz for balance-aware body execution.
Go to primary source ↓E11PDF p. 10, Table 3 and suite aggregation paragraph; p. 11, Sections 4.3.1–4.3.2
Twelve tasks have primary/overlap tags; every tagged task enters that suite average. Black-box evaluation uses preset scenes, randomized endpoints and order, hidden policy identity, fixed binary criteria, and ordinarily 20 trials/task/method.
Go to primary source ↓E12PDF p. 12, Figure 6, all five groups and legend; pp. 11–13, Section 4.3.3
Being-H0.7 leads Dynamic/Physical/Motion/Long Horizon/Generalization at 70.0/66.9/67.5/66.7/67.5%. H0.5: 47.5/59.4/46.7/60.6/57.5; π0.5: 49.2/55.0/45.8/56.1/58.3; Fast-WAM: 58.3/46.9/57.5/45.0/49.2. No uncertainty intervals or per-task success counts accompany these bars.
Go to primary source ↓E13PDF p. 12, Figure 7; p. 13, Visualization of the Latent Reasoning paragraphs
Six qualitative examples are shown. A video generator takes current observations and prior hidden states to synthesize future task states. No predictive-accuracy metric, observation-only control, generator implementation, or explicit border-color legend is supplied.
Go to primary source ↓E14PDF p. 13, Figure 8, both panels and legend; pp. 13–14, inference-infrastructure discussion
Latency/memory pairs: H0.5-UAC 3.45/5.6, H0.7-UAC 3.61/6.8, π0.5-RTC 12.45/9.3, H0.5 17.12/5.6, H0.7 28.81/6.8, Fast-WAM 41.26/15.6 (ms/step, GB). UAC locks committed actions and updates a future suffix using estimated delay. GPU model and precise latency definition are not given.
Go to primary source ↓E15PDF pp. 8–14, Section 4, Table 1 and Figures 4–8; p. 22, Appendix
Experiments comprise benchmark comparisons, physical evaluation, qualitative latent visualization, and deployment costs. No component ablation is reported; the appendix contains only author credits.
Go to primary source ↓E16PDF pp. 5–8, Sections 3–4.1; p. 13, deployment description; p. 22, Appendix
Some model/loss/data-format settings are specified, but not the full pretraining corpus recipe, optimizer schedule, duration/hardware, future sampling, norm threshold, rank-projection dimension, or denoising-step count. The appendix adds no implementation configuration.
Go to primary source ↓8.5 Primary sources
Being-H0.7: A Latent World-Action Model from Egocentric Videos ↗
PDF · 10,952 extracted words
Source fingerprint
e890143b3fd6564b9b1d43312d53bcb3a80000c7b79a090bb50506de0911f64d