HALO-WA: Hybrid-Attention Latent-Guided Online Reinforcement Learning for World-Action Models
1. Paper overview
In one sentence: A frozen world-action model supplies reference actions and visual memory to an online actor-critic, improving precision manipulation while depending on task-specific preparation and a competent base policy. e02e03e04e06e07e08e11e12
| At a glance | What to know |
|---|---|
| Research problem | Source description A plausible coarse trajectory can still miss an insertion or alignment by millimeters. The authors target deployment-specific perception, calibration, and contact errors while avoiding full online updates of a large WA backbone or learning control from scratch. e02 |
| Core mechanism | |
| A key reported result | Four real-world precision manipulation tasks: HALO-WA SR: 94.3 / 97.1 / 71.4 / 85.7, average 87.1. EL: 211.4 / 243.2 / 300.8 / 312.5, average 267.0. Success rate (%) and reported episode length (EL; counting convention unspecified). Task-specific ALOHA policies; 35 evaluation trials per task. Online training: Stick 45, Ethernet 60, Block 75, Power Plug 60 minutes. WA-base SR: 34.3 / 40.0 / 11.4 / 20.0, average 26.4; average EL 327.7. Average SR: Probe-Learn-Distill 52.1, HG-DAgger 62.9, RL-token-like 67.9. The reported average gain over RL-token-like is 19.2 percentage points. This establishes performance on four adapted tasks; no confidence intervals or repeated-training-seed variability are reported. e06e08e14 |
| Reading caution | Source description The authors restrict evidence to precision contact tasks on a limited platform set. Correction depends on useful base actions and latents; force, tactile feedback, and object dynamics are not explicitly modeled. e12 |
Core contributions
Figure 2. Reference actions organize the query sequence; visual latents provide the memory used to refine it. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow the diagram from left to right. The frozen backbone consumes robot state, the task prompt, and head/wrist images. Its outputs supply two distinct adapter inputs: reference-action tokens and visual tokens. In the upper actor branch, the robot-state and reference-action queries pass through self-attention before cross-attention reads visual keys and values. The output is the refined chunk sent to the environment. In the lower critic branch, candidate-action tokens join the queries; the fused tokens are mean-pooled into twin Q heads. The self-attention-to-cross-attention arrows agree with Section 3.2. Section 7.8 instantiates the output as twelve actions with fourteen dimensions each. e03e04e05e17
What it supports. The adapter can condition corrections on information beyond the final WA action. Its design also separates predicting an action chunk from evaluating that chunk: the actor drives execution, while the critic supports learning. This is the architectural basis for the report’s Dual-system assessment.
Where the evidence stops. The picture does not specify the exact latent extraction layer, tokenization, attention depth, or masks. The prose alternates between VAE/visual and action-generation latents; their precise implementation cannot be recovered from this diagram alone.
2. Motivation
2.1 The problem and the proposed response
A plausible coarse trajectory can still miss an insertion or alignment by millimeters. The authors target deployment-specific perception, calibration, and contact errors while avoiding full online updates of a large WA backbone or learning control from scratch. e02
2.2 What this reading follows
A robot can approach the right port and still miss the final insertion. HALO-WA treats this as an adaptation problem: retain the large world-action model, expose its internal visual latents, and train a smaller policy to correct its action chunks through interaction. The key architectural choice is to let state and reference-action tokens organize themselves before reading visual memory. This edition follows that information flow, then separates physical task performance, intervention behavior, and the narrower simulation evidence. The reported online minutes sit on top of demonstration collection and task-specific offline training; they are not the complete cost of obtaining the policy. e02e03e04e06e07e08e11e12
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 | WAMs |
| Architecture | Dual-system |
| Prediction paradigm | Other mechanisms |
| Quadrant | Outside quadrants |
3.1 Evidence-based assessment
Supports the recorded classification
The frozen WA generator and separately trained actor-critic support Dual-system. The added mechanism is latent-conditioned action refinement through online RL, not joint future/action prediction or inverse dynamics, supporting Other mechanisms and Outside quadrants. Policy post-training is demonstrated; broader generalization in the catalog subcategory remains untested. e03e04e05e12
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 1. Give the adapter both a proposal and evidence for revising it
Start with the distinction between the narrow robot state and the full RL state. Proprioception alone tells the adapter where the robot is, while the WA reference chunk tells it what the pretrained policy intends to do. The visual latent memory adds information that the authors argue can be lost when only final actions are exposed. HALO-WA puts state and reference actions on the query side, first models their relationships with self-attention, and then retrieves visual memory through cross-attention. The critic receives the same context plus a candidate chunk so that its value estimate depends on the proposed correction. My interpretation is that this interface makes correction conditional on the current proposal; the paper’s ablations support its practical usefulness, without proving which latent spatial cues cause the gain. e03e04e10e15e17
5.2 2. Learn from the action that actually moved the robot
The online loop must keep action labels consistent with physical transitions. When a human takes over, HALO-WA records the executed intervention actions rather than the actor’s displaced proposal. Section 7.1 additionally re-runs the WA model after intervention to complete the remaining chunk steps. The critic then learns from chunk-level reward and the next chunk boundary; the printed target uses gamma raised to the chunk length. The actor balances that critic’s value against BC toward reference behavior. Real-world demonstrations seed both actor warm-up and later mixed batches, helping explain why the short online phase does not begin from scratch. The source leaves important choices unresolved: how BC references are selected, how interrupted or terminal chunks are handled, and why the target equation omits the terminal flag mentioned in its prose. e05e07e13e18
5.3 3. Separate task improvement from a mechanism explanation
Read Table 1 as a comparison of adapted policies on four named physical tasks. It shows a strong average improvement, but the explanation that distributed latents preserve indispensable local information needs the additional ablations. Table 2 helps: removing latent guidance or changing the fusion method lowers success on Stick Insertion. Yet the supplement also says self-attention did not fully converge, so architecture and optimization are not cleanly separated. Figure 4 adds evidence about declining human takeover during training, not a replacement success metric. Finally, Table 3 changes the protocol by removing intervention and warm-up while retaining demonstrations. My assessment is that these results jointly support effective task-specific online correction, with much weaker evidence for a uniquely superior representation, standardized efficiency across methods, or generalization beyond the tested tasks. e06e07e08e09e10e11e12e15e16e18
5.4 Training and inference
During training
For each real task, 100 demonstrations train a separate GigaWorld Policy for 30,000 SFT steps. They also initialize the actor and enter a successful-demonstration buffer. Online updates mix demonstrations and replay equally while keeping the backbone frozen. These preparation costs precede the reported online minutes. e07e18
TD3 learns from executed chunks and sparse completion rewards. Human overrides replace proposed actions in replay. BC references may be WA actions, successful executed actions, or interventions; the paper does not specify a complete selection rule. Simulation randomly initializes the actor without warm-up or human intervention, but retains demonstration sampling. e05e16e18
During inference
Cloud computation runs the WA backbone and actor; the robot executes the refined chunk and returns observations. Online learning is asynchronous. After an intervention, WA inference is repeated from the current observation to fill the remaining chunk steps. The described control path does not use critic-guided search through imagined trajectories. e03e05e13
5.5 Implementation flow
- Extract the prior and memory
The backbone generates H reference actions; the adapter uses the first C, with C ≤ H. The RL state includes proprioception, this reference chunk, and WA latents. Thus the action prior and latent memory both remain available at deployment. e03
- Refine through two attention operations
Robot-state and reference-action tokens first undergo self-attention, then query visual tokens as cross-attention keys/values. The actor decodes action tokens into a complete refined chunk; no explicit residual-addition equation or inverse-dynamics stage is specified. e03e04
- Evaluate candidate chunks
The critic adds candidate-action tokens to the query side, applies the same attention organization, mean-pools the fused representation, and uses two Q heads. The implementation has 13 actor queries, 25 critic queries, hidden width 512, and twelve 14-dimensional output actions. e04e17
6. Experiments & results
HALO-WA uses a frozen world-action policy twice: its action chunk supplies a behavioral prior, and its distributed visual latents supply information for correction. A separate actor-critic learns refined chunks from real interaction. Four physical tasks show strong task-specific gains, supported by narrower ablations and simulation experiments; general-purpose transfer is untested.
6.1 Read the original evidence
Table 1. The four adapted physical tasks improve, with Block Assembly remaining the least successful. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the success-rate block first, comparing methods within each task column. The last row gives HALO-WA’s result; the preceding RL-token-like row is the strongest baseline on average, although HG-DAgger is stronger on Stick Insertion. Then move to the episode-length block, where smaller values are reported as preferable. The parenthesized changes under HALO-WA are relative changes against WA-base, not percentage-point gains. Section 4.1 reports 35 evaluation trials per task. Section 7.2 explains that each task starts from a separately fine-tuned base policy trained on 100 demonstrations, which also support actor initialization and online sampling. e06e07e08e12e19
What it supports. Average success rises from 26.4% for WA-base to 87.1% for HALO-WA, exceeding RL-token-like’s 67.9% by the reported 19.2 percentage points. The per-task HALO-WA rates range from 71.4% on Block Assembly to 97.1% on Ethernet Insertion. These are physical execution outcomes after task-specific adaptation.
Where the evidence stops. No confidence intervals or training-seed variability are reported. EL aggregation and units need clarification: Stick Insertion EL is 211.4 despite a listed training horizon of 192. The table does not establish transfer to unseen tasks.
Figure 4. Human takeover becomes less frequent during learning, but progress is not monotonic on every task. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Use the shared legend above the panels: blue is HALO-WA, purple dashed is RL-token-like, and orange is HG-DAgger. The vertical axis is intervention rate in percent, while the horizontal axis is training time in minutes. Read each panel within its own time range rather than comparing equal horizontal positions across tasks. Stick Insertion declines relatively quickly; Block Assembly first improves, then rises again before approaching zero. This matters because the plot summarizes a learning process rather than a fixed-policy test. The caption’s late-stage claim can be checked directly, while the exact counting window for intervention rate is not supplied. e06e09e12e13
What it supports. HALO-WA’s curves support reduced dependence on human corrections late in online training across these four tasks. They complement the final success table by showing how takeover behavior changes during adaptation. The Block Assembly rebound qualifies any suggestion that improvement is uniformly steady throughout learning.
Where the evidence stops. Intervention rate is not a force-safety metric or a success probability. Operator thresholds, counting rules, smoothing, and run-to-run uncertainty are not documented; near-zero takeover should not be interpreted as a general safety guarantee.
Table 3. Simulation shows improvement without human takeover, with a substantial remaining gap on the hammer task. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Compare the two methods within each task block. Section 4.4 evaluates each method on 100 trials, so the success rates are percentages of that task’s trials. The parenthesized values are relative changes against WA-base; the large relative gain on Beat Block Hammer coexists with only 45% absolute success. The surrounding text and Section 7.7 distinguish this setup from the physical experiments: the actor is randomly initialized and receives no human intervention or WA-behavior warm-up. However, Section 7.9 still uses a demonstration buffer with half of each training batch drawn from demonstrations. This is not learning from interaction alone. e11e16e18e19
What it supports. HALO-WA improves Click Bell from 54% to 97% and Beat Block Hammer from 22% to 45%. The result supports online adaptation without human overrides in these two simulated tasks. It also shows that removing intervention does not yield uniformly high success across tasks of different complexity.
Where the evidence stops. The simulation data count and online training budget are unspecified. Both tasks list a 192-step training horizon, yet hammer EL exceeds 192 for both methods. Different evaluation horizons or counting conventions remain possible but unexplained.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Four real-world precision manipulation tasks Task-specific ALOHA policies; 35 evaluation trials per task. Online training: Stick 45, Ethernet 60, Block 75, Power Plug 60 minutes. | HALO-WA SR: 94.3 / 97.1 / 71.4 / 85.7, average 87.1. EL: 211.4 / 243.2 / 300.8 / 312.5, average 267.0. Success rate (%) and reported episode length (EL; counting convention unspecified) | WA-base SR: 34.3 / 40.0 / 11.4 / 20.0, average 26.4; average EL 327.7. Average SR: Probe-Learn-Distill 52.1, HG-DAgger 62.9, RL-token-like 67.9. The reported average gain over RL-token-like is 19.2 percentage points. This establishes performance on four adapted tasks; no confidence intervals or repeated-training-seed variability are reported. e06e08e14 |
| Stick Insertion mechanism ablations Real-world Stick Insertion; Section 4.3 and variant definitions in Sections 7.5–7.6. | Hybrid attention / WA latent features: 94.3. Success rate (%) | Direct concatenation 28.6; CNN 74.3; cross attention 57.1; self attention 62.9. Raw-image encoder 25.7; WA intermediate features 57.1; no latent 22.9. The ranking supports the chosen interface within this task and budget. Parameter counts and matched compute are absent; self-attention is explicitly described as not fully converged. e10e15 |
| RoboTwin Click Bell and Beat Block Hammer 100 evaluation trials per method per task; randomized initial conditions; no human intervention or actor warm-up, with demonstrations still sampled. | Click Bell: 97.0 SR, 100.3 EL. Beat Block Hammer: 45.0 SR, 213.7 EL. Success rate (%) and reported EL | WA-base: Click Bell 54.0 SR, 151.4 EL; Beat Block Hammer 22.0 SR, 312.5 EL. Online improvement occurs without human takeovers, but hammer success remains below half. EL cannot be reconciled with the listed 192-step training horizon without further protocol details. e11e16e18e19 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 2. Both the fusion operation and the representation supplied to it matter in the reported Stick Insertion experiment. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Treat the two italicized blocks as separate comparisons. The upper block changes how latent, state, and action information are fused: direct concatenation, CNN compression, cross-attention, self-attention, or the proposed hybrid. The lower block changes the visual representation: a raw-image encoder, WA intermediate features, no latent input, or WA latent features. The repeated 94.3/211.4 row is the full method appearing in both comparisons, not a second independent result. Read success rate upward and episode length downward as indicated by the headers. Sections 7.5–7.6 are essential companions because they define these variants and describe their differing convergence behavior. e10e15
What it supports. The full method reaches 94.3% success versus 74.3% for CNN fusion, 62.9% for self-attention, and 57.1% for cross-attention. Removing latent input yields 22.9%. This supports the usefulness of the selected latent interface under the tested setting, while leaving its causal advantage over equally trained alternatives open.
Where the evidence stops. These ablations cover one physical task. The authors say self-attention did not fully converge within the available training budget. Without parameter counts, matched compute, or repeated seeds, the table cannot isolate architecture from optimization differences.
7. Analysis & limitations
7.1 What the evidence leaves open
The authors restrict evidence to precision contact tasks on a limited platform set. Correction depends on useful base actions and latents; force, tactile feedback, and object dynamics are not explicitly modeled. e12
Figure 4 shows decreasing intervention, with nonmonotonic behavior on Block Assembly. It does not measure force safety or independent generalization; intervention counting and uncertainty are not specified. e09
The tables report point estimates without confidence intervals. Train/test object separation, baseline tuning and compute parity, and EL aggregation are insufficiently documented. Stick EL 211.4 exceeds its listed 192-step training horizon; hammer EL has the same issue. Different evaluation horizons or units could explain this, but are not stated. e06e08e10e11e15e19
7.2 Questions for discussion
- Would hybrid attention still outperform self-attention after matching parameter count, update budget, and convergence?
- How much improvement remains when the WA prior is systematically misaligned, rather than already broadly competent?
- Which terminal and episode-length conventions reconcile the objectives, training horizons, and reported EL?
8. Reproducibility audit
8.1 Requirements and known gaps
Required setup includes the task-specific WA checkpoints, demonstration data, an ALOHA robot with three 256×256 RGB views, and human takeover recording. Real-world compute uses two A800 GPUs; Click Bell specifies one A800 training GPU. The paper calls the robot Songling in Section 4.1 and AgileX in Section 7.1. e06e07e11e13
Reported settings are UTD 5, critic/actor update ratio 2, BC coefficient 50, batch 64, micro-batch 4, and actor/critic learning rates 2.5×10⁻⁵ / 3.0×10⁻⁵. Table 4 gives (horizon, γ): Stick (192, .92), Plug/Ethernet (384, .95), Block (480, .98); simulation uses 192 with .82 for Bell and .92 for Hammer. e18e19
Unresolved implementation inputs include the exact WA checkpoint/configuration, latent extraction layer and tokenization, attention depth/heads, action normalization, TD3 noise and target-update settings, terminal/truncation handling, warm-up duration, simulation data count and training budget, and software versions. Sparse positive-reward magnitude is unspecified. Resolve these before calling any run a faithful reproduction. e05e07e16e17e18e19
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: Match the attention comparison before attributing the gain
Reader-proposed experiment, not performed: on Stick Insertion, compare hybrid attention, self-attention, and CNN fusion using the same frozen task-specific WA checkpoint, latent tensors, demonstrations, action chunking, BC settings, and a predeclared human-intervention policy. Match trainable parameter counts as closely as feasible and report both equal-interaction/equal-update results and extended-training results. Repeat independent training seeds and evaluate on a shared set of held-out resets without takeover. Track success, intervention during training, updates, and wall-clock cost. If self-attention catches up after convergence, the published ranking would support faster optimization under that budget more strongly than a uniquely better representation; a persistent hybrid advantage would strengthen the architectural claim. e06e07e10e15e17e18
Check 2: Audit chunk boundaries, terminal targets, and EL in Click Bell
Reader-proposed implementation check, not performed: build a small Click Bell run using the reported 12-action chunks, 192-step training horizon, gamma 0.82, sparse rewards, and 1:1 demonstration/replay sampling. Log primitive steps, chunks, terminal success, time-limit truncation, accumulated chunk reward, and bootstrap values. Compare a literal implementation of Eq. (5) with an explicitly documented terminal-masked variant, holding all other choices fixed; report evaluation horizons separately from training horizons. Check terminal examples by hand, then compare critic targets and success across repeated seeds. A nonzero terminal bootstrap in the literal variant exposes the missing convention; differing EL under step/chunk counting can test possible explanations for the paper’s horizon discrepancies. Neither convention should be claimed as the authors’ implementation without clarification. e05e11e16e18e19
8.3 Reading coverage
Visual audit: The title, all ten authors, affiliation block, and v1 date were inspected on p. 1. Visual reading covered the overview on p. 2, method framing on p. 4, Figure 2 and Eqs. (1)–(3) on p. 5, attention and Eqs. (4)–(7) on p. 6, setup and Figure 3 on p. 7, Tables 1–2 and Figure 4 on p. 8, and simulation results and limitations on p. 9. Appendix pages 12–17 were visually inspected for offline data, intervention handling, task sequences, baseline/ablation definitions, simulation protocol, adapter dimensions, rewards, hardware-related context, and hyperparameter tables. All five final crops were separately viewed; the real-world table crop was tightened and re-viewed. Figure 2 query/key/value roles and self-to-cross attention arrows agree with Section 3.2; Eq. (5)'s absent terminal mask and the result-EL/training-horizon discrepancies remain disclosed. Page 3 and reference pages 10–11 were text-read only. Separate supplements, code, and external resources were not inspected.
PDF pages inspected for this edition: 1, 2, 4, 5, 6, 7, 8, 9, 12, 13, 14, 15, 16, 17. 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
- Title, author block, abstract (p. 1)
- 1 Introduction (pp. 1–4)
- 2 Related Work, 2.1–2.2 (p. 4)
- 3 Method, 3.1–3.3 (pp. 4–6)
- 4 Experiments, 4.1–4.4 (pp. 7–9)
- 5 Limitations and 6 Conclusion (p. 9)
- References (pp. 10–11)
- 7 Supplementary Material: 7.1 implementation and 7.2 baselines (pp. 12–13)
- 7.3 task details and 7.4 additional results analysis (pp. 13–14)
- 7.5 ablation variants and 7.6 discussion (pp. 14–15)
- 7.7 simulation implementation (pp. 15–16)
- 7.8 adapter architecture and 7.9 hyperparameters (pp. 16–17)
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- The inspected title page identifies arXiv:2607.04265v1 [cs.RO], 5 July 2026. Its title and all ten authors agree with the catalog after name-order normalization. No other revision was supplied or compared. The title page additionally lists affiliation 3, GigaAI, Beijing, China, omitted from the catalog affiliation string.
- Text extraction does not reconstruct figure images; the retained PDF was separately inspected for figures and equation/table layout. All five supplied text chunks were read individually, including the complete embedded supplementary material.
- Separate supplemental material availability has not been fully verified.
- Code, linked resources, and external papers were not inspected; no experiments were reproduced. The code-release statement is an author statement, not verified repository availability.
- Visual inspection covered PDF pages 1, 2, 4–9, and 12–17. Page 3 and reference pages 10–11 were read as text only. The main text calls the supplement Appendix A–G, whereas the supplied PDF labels it Sections 7.1–7.9.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title/author/affiliation block and arXiv margin stamp
The observed title is HALO-WA: Hybrid-Attention Latent-Guided Online Reinforcement Learning for World-Action Models; arXiv:2607.04265v1, 5 July 2026. Authors in order: Angen Ye, Weijie Ke, Xiaofeng Wang, Xinze Chen, Chaojun Ni, Guosheng Zhao, Boyuan Wang, Zheng Zhu, Junjie Xie, Dapeng Zhang. Four affiliations include GigaAI. Ye and Ke share equal contribution; Zhu is corresponding author.
Go to primary source ↓e02PDF pp. 1–2, Abstract and Section 1
Motivates end-stage calibration, perception, and contact errors; proposes retaining pretrained priors while adapting a small policy through online interaction.
Go to primary source ↓e03PDF p. 2, Figure 1 and final two paragraphs; p. 5, Figure 2 and Section 3.1, Eqs. (1)–(3)
Cloud inference and learning feed chunks to a robot execution loop. Frozen WA produces reference actions and latents; the adapter state combines robot state, the first C of H reference actions, and WA latents, then directly predicts a C-by-A refined action chunk.
Go to primary source ↓e04PDF pp. 5–6, Figure 2, caption, and Section 3.2
Actor query tokens contain state and reference actions; critic additionally receives candidate actions. Self-attention precedes cross-attention, with visual memory as keys/values. Twin critic heads estimate values.
Go to primary source ↓e05PDF p. 6, Section 3.3, Eqs. (4)–(7) and surrounding definitions
Replay stores executed chunks, rewards, next states, and terminal flags. Target uses accumulated reward plus gamma to the C power times minimum target-Q, without a printed terminal mask. Actor optimizes Q plus BC; possible BC targets include WA, successful executed, and intervention actions.
Go to primary source ↓e06PDF p. 7, Section 4.1 and Figure 3
Songling ALOHA, three RGB cameras, two A800 GPUs, actor warm-up, 35 trials per real task, and online times 45/60/75/60 minutes. Baselines and SR/EL/intervention metrics are named.
Go to primary source ↓e07PDF p. 12, Section 7.2, WA-base paragraphs; p. 13, HALO-WA paragraph
Separate task-specific GigaWorld Policies use 100 demonstrations per task and 30k SFT steps. The demonstrations also warm up the actor and enter the demonstration buffer, sampled equally with online replay. RL-token-like uses a compact WA-latent representation.
Go to primary source ↓e08PDF p. 8, Table 1, all method rows and SR/EL columns; Section 4.2 continuation
HALO-WA SR values are 94.3/97.1/71.4/85.7, average 87.1, versus WA-base 34.3/40.0/11.4/20.0, average 26.4. Other average SRs: 52.1, 62.9, 67.9. HALO-WA EL values are 211.4/243.2/300.8/312.5, average 267.0, versus WA-base average 327.7. No uncertainty accompanies the table.
Go to primary source ↓e09PDF p. 8, Figure 4, legend, axes, and caption
Intervention rate (%) is plotted against training minutes for HG-DAgger, RL-token-like, and HALO-WA on four tasks. HALO-WA approaches zero late; the Block Assembly curve includes a rebound. No uncertainty bands or operational counting definition are supplied.
Go to primary source ↓e10PDF p. 8, Table 2, Hybrid-Attention and Latent-Guided blocks; Section 4.3
Stick Insertion SR/EL: concat 28.6/276.4, CNN 74.3/229.7, cross 57.1/246.9, self 62.9/239.1, hybrid 94.3/211.4; raw image 25.7/282.6, intermediate 57.1/248.3, no latent 22.9/287.5, WA latent 94.3/211.4.
Go to primary source ↓e11PDF p. 9, Section 4.4 and Table 3, both task blocks
100 evaluation trials per method per task; no intervention or actor warm-up. Bell improves SR/EL from 54.0/151.4 to 97.0/100.3, Hammer from 22.0/312.5 to 45.0/213.7. Bell uses one A800 training GPU. Parenthesized percentages are relative changes.
Go to primary source ↓e12PDF p. 9, Section 5 Limitations
Limited task/platform coverage, dependence on useful WA priors/latents, and no explicit contact-force, tactile, or object-dynamics modeling are acknowledged.
Go to primary source ↓e13PDF p. 12, Section 7.1
Names AgileX ALOHA, resizes three camera images to 256×256, records human-executed actions, refills remaining chunk steps by WA reinference after interventions, and asynchronously trains only the adapter.
Go to primary source ↓e14PDF pp. 13–14, Section 7.3, four task paragraphs and unnumbered image sequences
Stick requires axial alignment, Ethernet connector alignment and stable insertion, Block geometric assembly, and Plug orientation/localization/contact. Online times are 45, 60, 75, and 60 minutes respectively.
Go to primary source ↓e15PDF pp. 14–15, Sections 7.5–7.6; p. 14, Section 7.4
Defines concatenation, CNN, cross/self/hybrid attention, raw-image, intermediate-feature, and no-latent variants. States self-attention has higher training cost and does not fully converge within the real-world budget. Compression explanations are author interpretations; parameter/compute matching is not documented.
Go to primary source ↓e16PDF pp. 15–16, Section 7.7
RoboTwin tasks randomize object positions and initial states, use few-shot WA-base training, random actor initialization, no human intervention, and asynchronous rollout/training. Counts and online budgets are not specified.
Go to primary source ↓e17PDF p. 16, Section 7.8
Hidden dimension 512; 13 actor query tokens, 25 critic query tokens; actor MLP 512→1024→512→14 yields 12×14 actions; mean-pooled critic feeds independent 512→2048→1024→512→1 heads. Exact attention depth and latent-token construction are not given.
Go to primary source ↓e18PDF p. 16, Section 7.9
All tasks use 12×14 chunks, 1:1 demo/replay batches, sparse positive success reward and zero otherwise. Gives UTD 5, critic/actor ratio 2, BC 50, batch 64, micro-batch 4, actor LR 2.5×10⁻⁵, critic LR 3.0×10⁻⁵; positive reward magnitude is unspecified.
Go to primary source ↓e19PDF p. 17, Tables 4–5 and explanatory paragraphs
Real (maximum episode length, gamma): Stick (192,.92), Plug/Ethernet (384,.95), Block (480,.98). Simulation: Aloha-AgileX, both horizons 192, Bell gamma .82 and Hammer .92. Tables do not explain how these horizons relate to EL in result tables.
Go to primary source ↓8.5 Primary sources
HALO-WA: Hybrid-Attention Latent-Guided Online Reinforcement Learning for World-Action Models ↗
PDF · 8,516 extracted words
Source fingerprint
aa1aed4da00ee57b3274d44a26ff966647a2736dab01144e3062c56f0ffdd50e