Helix: A Vision-Language-Action Model for Generalist Humanoid Control
1. Paper overview
In one sentence: Helix jointly learns a slow semantic conditioner and a fast upper-body control policy, with latency-aware training but no controlled evaluation isolating that design's benefits. e-identitye-probleme-duale-architecturee-architecture-visuale-traininge-inferencee-upper-bodye-collaboratione-grasping
| At a glance | What to know |
|---|---|
| Research problem | Author claim Household manipulation requires responding to unfamiliar objects and instructions without programming or collecting demonstrations for every behavior. The overview frames the obstacle as a speed–generality tradeoff: pretrained vision-language models offer semantic breadth, while fast visuomotor policies supply responsive control. Helix attempts to combine these properties for upper-body humanoid manipulation. e-probleme-dual |
| Core mechanism | Source description Two jointly trained systems communicate through one continuous semantic vector: System 2 supplies task conditioning, and System 1 produces high-rate continuous actions. Their distinct architectures and update rates are explicit. e-duale-architecture |
| A key reported result | Whole-upper-body humanoid control: 35 degrees of freedom at 200 Hz; source-reported operating characteristics. Reported action-space dimensionality and control frequency. Figure’s robot demonstration described under Results and Video 3; no named benchmark or trial protocol. The source claims no prior VLA demonstrated this combination of coordination and generalization; it provides no matched numerical baseline. The account describes coordinated fingers, wrists, head and torso. These are reported control dimensions and frequency, not manipulation success, measured latency or new-image frequency. Trial counts and outcome uncertainty are absent. e-upper-bodye-architecture-visual |
| Reading caution | Reader analysis Results are narrative accounts associated with videos. No quantitative outcome table, confidence interval, controlled ablation or systematic failure analysis is provided. The causal contributions of semantic pretraining, latent conditioning and latency-matched training therefore remain unresolved by this overview. e-upper-bodye-collaboratione-graspinge-traininge-designe-discussion |
Core contributions
- Source description
Two jointly trained systems communicate through one continuous semantic vector: System 2 supplies task conditioning, and System 1 produces high-rate continuous actions. Their distinct architectures and update rates are explicit. e-duale-architecture
- Author claim
Figure reports that one shared set of Helix weights supports picking, container interaction and robot handovers without task-specific fine-tuning. Claims of priority and commercial readiness are publisher assertions, not independently established comparisons. e-introductione-traininge-discussion
- Author claim
Figure claims the speed of specialized single-task behavioral-cloning policies together with zero-shot handling of thousands of test objects. It contrasts continuous outputs with action tokenization used for lower-dimensional setups such as binarized parallel grippers, and argues that separate S1/S2 components simplify iteration. These are source-specific design comparisons; no matched speed/generalization or tokenization ablation is reported. e-designe-inference
Original Figure 1 and adjacent text. A conceptual argument for specifying new behaviors through language, with no measured scaling curve. Original source, original section ↗
Extracted original source content with a documented offline layout wrapper and local Arial font; not a historical website screenshot or publisher PDF. Original Figure 1 raster and original adjacent text/caption; conceptual illustration, not a measured scaling-law experiment.
How to read it. Start with the axes: Number of tasks increases vertically and Time runs horizontally, but neither has numerical ticks. Then read the three panels on the right. The traditional approach associates behavior growth with PhD hours; the fleet model associates it with collected data; the Helix panel claims new behaviors without programming or demonstrations. The purple curve illustrates that ambition rather than reporting sampled performance. Read the retained caption together with the training account: Helix still uses approximately 500 hours of robot demonstrations and pretrained components. The useful distinction is between learning the system and specifying another behavior after learning. e-probleme-datae-architecturee-traininge-conclusion
What it supports. The figure explains why the source prioritizes language-conditioned generalization: manually authoring or demonstrating every household behavior would be expensive. It motivates the S2-to-S1 interface, but provides no measured task-growth rate, cost reduction or speedup. The source's separate training account documents the robot supervision and component pretraining actually used.
Where the evidence stops. The slogan about behaviors without training must be scoped to new behavior specification; it does not erase robot supervision or pretraining. Uncalibrated curves cannot establish a scaling law, and the future 1,000-fold scaling ambition in the conclusion is unevaluated.
2. Motivation
2.1 The problem and the proposed response
Household manipulation requires responding to unfamiliar objects and instructions without programming or collecting demonstrations for every behavior. The overview frames the obstacle as a speed–generality tradeoff: pretrained vision-language models offer semantic breadth, while fast visuomotor policies supply responsive control. Helix attempts to combine these properties for upper-body humanoid manipulation. e-probleme-dual
2.2 What this reading follows
Figure's February 20, 2025 overview asks how language understanding can guide dexterous household manipulation without collecting demonstrations for every new behavior. Helix connects a pretrained vision-language model, System 2, to a reactive transformer, System 1, through one continuous latent vector. The original architecture distinguishes 20 Hz sensory branches, 7–9 Hz semantic updates and 200 Hz action outputs. The tutorial follows that interface from supervised training to asynchronous deployment, then separates reported control capabilities from demonstrated reliability. Two original visuals supply the motivation and mechanism; the results remain textual accounts of upper-body control, shared-weight collaboration and novel-object grasping, without success-rate denominators or controlled ablations. e-identitye-probleme-duale-architecturee-architecture-visuale-traininge-inferencee-upper-bodye-collaboratione-grasping
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 | VLA |
| Architecture | Dual-system |
| Prediction paradigm | Not applicable |
| Quadrant | Not applicable |
3.1 Evidence-based assessment
Supports the recorded classification
The recorded hierarchical/dual-system VLA classification is supported by the explicit S2-to-S1 latent interface, different model architectures and asynchronous execution. A single shared set of task weights and joint optimization do not make this a single architectural system. The supplied mechanism maps observations and language to control; it specifies no joint future/action prediction, inverse-dynamics extraction or inference-time world rollout. The catalog’s prediction-paradigm and quadrant entries are therefore appropriately not applicable within this overview’s scope. e-duale-architecturee-traininge-inference
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 |
|---|---|
|
|
5. Method in detail
5.1 Learn a reusable language-to-control interface
The scaling figure motivates changing how a user specifies behavior after training. The actual learning pipeline starts from approximately 500 hours of multi-robot, multi-operator teleoperation. A labeling VLM observes segmented onboard-camera clips and supplies hindsight instructions describing the demonstrated behavior. S2 combines the command, monocular images and robot state into a continuous semantic vector. S1 projects that vector into its token space and concatenates it with its own visual features. Action regression sends gradients through the vector from S1 into S2. The source calls this one Helix training stage with no per-task heads or fine-tuning; it separately describes internet pretraining for S2 and simulation pretraining for S1's vision backbone. Shared behavior weights therefore coexist with two distinct architectural components. e-probleme-datae-architecturee-traininge-discussion
Original architecture animation-derived still and adjacent source section. Semantic conditioning and reactive control run at different rates; this component-timed original animation still is not a continuous-playback record. Original source, original section ↗
Extracted original source content with a documented offline layout wrapper and local Arial font; not a historical website screenshot or publisher PDF. Derived architecture still from the original desktop animation: Lottie frame 200 at 100 fps (2 s), each embedded GIF independently at frame 0 (0 ms), source-parent #0C0C0C wrapper matte. Outlined original labels have no font dependency. This component-time tuple is not synchronized historical playback or full-video inspection. The diagram says Joint angles; nearby prose describes wrist pose, fingers, and upper body. Inputs are 20 Hz, S2 runs 7–9 Hz, and outputs are 200 Hz.
How to read it. Follow the command into System 2, the 7B pretrained VLM labeled GPU 2. Its Latent Vector passes to System 1, the 80M transformer labeled GPU 1, at 7–9 Hz. Trace the parallel image and state branches: the S2 branches read 7–9 Hz, while the branches into S1 read 20 Hz. The rightward whole-upper-body output reads 200 Hz. The original arrows therefore separate sensory input, semantic updates and action output. Cross-check state labels before implementing: the graphic says Joint angles and Finger positions, whereas the Architecture prose specifies wrist pose and finger positions. No equivalence or conversion between those descriptions is supplied. e-duale-architecturee-architecture-visuale-actionse-inference
What it supports. The graphic supports a dual-system architecture with a learned semantic interface and a faster action loop. The adjacent deployment account explains asynchronous shared-memory conditioning on separate GPUs. A 200 Hz output does not establish 200 Hz fresh camera input, and the still's robot imagery does not verify successful continuous collaboration.
Where the evidence stops. Only the disclosed outer frame at 2 s and independent GIF frame-zero components were viewed, not synchronized playback. Joint-angle versus wrist-pose naming and the detailed schedule between 20 Hz inputs and 200 Hz outputs remain unresolved; no interpolation mechanism is inferred.
5.2 Match semantic delay to the deployed control loop
During training, the source offsets S1 and S2 inputs to reflect their deployed latency gap. At deployment, S2 runs in the background on its own GPU, updating a shared latent at 7–9 Hz; S1 uses the latest available latent and observation in a 200 Hz control loop on another GPU. The diagram's S1 sensory branches are labeled 20 Hz, so repeated action outputs must not be described as independently observed camera frames. Reader interpretation: latency matching could make the motor policy less sensitive to stale intent, but the overview supplies no ablation establishing that benefit. It also leaves the observation schedule and state-name discrepancy unresolved. This explicit S2-to-S1 interface supports the recorded dual-system VLA classification; the described deployment contains no future-world rollout or search procedure. e-traininge-inferencee-duale-architecture-visuale-architecture
5.3 Separate breadth claims from measured reliability
The Results section reports a 35-DoF upper-body action space at 200 Hz, two robots storing unseen groceries with identical weights, and thousands of novel household objects handled in clutter. These describe operating scope and claimed behavior breadth, without successful-trial denominators or uncertainty. The source says training objects are excluded from evaluation; reader interpretation is that this supports object-instance separation, not absence of those concepts from internet pretraining. The collaboration example also specifies which robot gives and receives, qualifying the claim of no explicit role assignments. Finally, approximately 500 hours being less than 5% of earlier VLA datasets has no named comparison denominator, and claimed specialist-policy speed has no matched timing table. Neither those comparisons nor the conceptual scaling curves establish a controlled efficiency gain. e-upper-bodye-collaboratione-graspinge-datae-architecturee-discussione-designe-inferencee-problem
5.4 Training and inference
During training
The dataset contains approximately 500 hours of diverse teleoperation from multiple robots and operators. An auto-labeling VLM examines segmented onboard-camera clips and generates hindsight instructions describing the demonstrated action. The source states that all objects handled in training are excluded from evaluation; it does not provide an object list or a detailed split protocol. e-data
Helix is trained end to end with a standard regression loss on continuous actions. Gradients flow from S1 into S2 through the conditioning vector. The Helix training stage uses shared weights without per-task heads or adaptation, after the stated component pretraining. A temporal offset between S1 and S2 training inputs is calibrated to their deployed latency gap. The exact loss formula and offset value are not given. e-traininge-architecture
During inference
Each robot has two embedded low-power GPUs, with S2 and S1 assigned dedicated GPUs. S2 asynchronously consumes the latest camera/state observation and instruction and refreshes a shared-memory latent at the reported 7–9 Hz rate. S1 uses the latest available observation and latent in its separate 200 Hz control loop. The source links this asynchronous schedule to the training offset and claims the speed of its fastest single-task imitation policies. GPU models, numerical power consumption and a matched latency benchmark are not supplied. e-inferencee-dual
The architecture explicitly labels S1 image/state inputs 20 Hz and outputs 200 Hz. A fast output loop therefore does not establish a new camera observation for every action. The prose's claim of higher temporal resolution is relative to S2; intermediate observation handling, interpolation or reuse is unspecified. Although the source calls S2's role latent planning, it describes a conditioning vector rather than a future-state rollout or search algorithm. e-architecture-visuale-inferencee-architecture
5.5 Implementation flow
- Encode semantic intent in System 2
S2 uses a 7B-parameter open-source, open-weight VLM pretrained on internet-scale data. Images and robot state are projected into the vision-language embedding space and combined with the command. Task-relevant information is distilled into a single continuous latent vector. The overview reports an S2 rate of 7–9 Hz. e-architecturee-dual
- Condition reactive perception in System 1
S1 is an 80M-parameter cross-attention encoder-decoder transformer. Its fully convolutional, multiscale vision backbone is initialized from simulation-only pretraining. The projected S2 vector is concatenated with visual features along the sequence dimension. S1 receives the same image/state input types as S2 and is described as processing them more frequently. The diagram labels the S1 image/state branches 20 Hz and the S2 branches 7–9 Hz; these input rates must not be equated with the 200 Hz action-output loop. e-architecturee-architecture-visual
- Produce control and completion targets
S1 outputs continuous upper-body targets at 200 Hz, covering wrists, fingers, torso and head. A synthetic percentage-task-completion action predicts termination to help sequence behaviors. The description does not specify the completion threshold or the lower-level actuator controller that executes these targets. e-actions
- Keep diagram and prose conventions distinct
The diagram names robot state Joint angles and Finger positions; the Architecture prose instead specifies wrist pose and finger positions. Output prose adds desired wrist poses, finger flexion/abduction, and torso/head orientation targets. These are different descriptions of input state and output targets, not an established equivalence. The overview gives no transform or detailed timing schedule resolving the representation difference or how the 20 Hz input branches feed the 200 Hz output loop. e-architecture-visuale-architecturee-actionse-inference
6. Experiments & results
Helix connects an internet-pretrained vision-language model to a faster visuomotor policy through a learned continuous latent vector. Figure's overview describes natural-language-conditioned upper-body control, including novel-object grasping and two-robot grocery storage. S2 supplies semantic updates at 7–9 Hz; S1's control loop outputs at 200 Hz, while the original architecture labels S1's image/state input branches at 20 Hz. Joint action-regression training includes a latency-matched input offset. The inspected figures explain the motivation and architecture; task outcomes remain publisher-reported accounts without success-rate denominators, controlled ablations or measured scaling curves.
This technical overview has two meaningful original mechanism/motivation visuals: the conceptual raster and the architecture animation represented by the explicitly derived still. It supplies no original quantitative results table, ablation figure or measured scaling experiment, so this edition has two visuals rather than four to six and no quantitative-table or results/ablation visual. The ten supporting text excerpts are evidence context, not extra invented figures. Other external image assets remain unavailable; external Videos 1–6 were not watched. Architecture GIFs and the full animation were not viewed continuously. The still supports visible labels and component layout only, not motion, latency measurement or robot success. The source's qualitative result claims are preserved as claims. e-probleme-architecture-visuale-mediae-upper-bodye-collaboratione-graspinge-traininge-discussion
6.1 Read the original evidence
No results visual is included in this edition. The source-specific visual limitations are recorded above.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Whole-upper-body humanoid control Figure’s robot demonstration described under Results and Video 3; no named benchmark or trial protocol. | 35 degrees of freedom at 200 Hz; source-reported operating characteristics. Reported action-space dimensionality and control frequency | The source claims no prior VLA demonstrated this combination of coordination and generalization; it provides no matched numerical baseline. The account describes coordinated fingers, wrists, head and torso. These are reported control dimensions and frequency, not manipulation success, measured latency or new-image frequency. Trial counts and outcome uncertainty are absent. e-upper-bodye-architecture-visual |
| Zero-shot collaborative grocery storage Two Figure robots use identical model weights, unseen groceries and natural-language handover/storage instructions. | Successful two-robot collaboration reported in the text; number of trials, successful/attempted-task denominator, success rate and uncertainty absent. Qualitative task demonstration | No controlled coordination or specialist-policy baseline. The source describes physical robot manipulation. Its same-weight deployment supports the reported collaboration setup, but its no-explicit-role-assignment claim is qualified by example instructions that specify giving, receiving, left and right. Autonomous role discovery is not established. e-collaboration |
| Language-conditioned grasping of novel household objects Cluttered scenes with objects excluded from robot training; direct object names and conceptual descriptions. | The source reports thousands of novel items handled in systematic testing, without an exact number of objects or attempted grasps, success rate, or uncertainty. Qualitative breadth of handled objects | No matched quantitative baseline or per-object breakdown. The reported toy-cactus response to a desert-item instruction illustrates claimed semantic transfer and hand selection. Robot-training object exclusion does not establish that semantic categories were absent from VLM pretraining, or that arbitrary household objects can reliably be grasped. e-graspinge-datae-architecture |
| Training-data scale comparison Discussion's resource comparison for Helix's supervised robot training, with pretrained S2 and S1 vision components described separately in Architecture. | Approximately 500 hours; the source describes this as <5% of previously collected VLA datasets. Reported teleoperation hours and stated fraction of earlier VLA datasets | No reference dataset is named, denominator total or normalization is supplied, and no matched performance/data-budget experiment is provided. The source compares the collection scale to modern single-task imitation datasets and claims no dependency on multi-robot-embodiment collection or multiple Helix training stages. This is a publisher-reported resource comparison, not a quantified sample-efficiency gain. Multi-robot collection is explicitly reported; it does not by itself establish multiple embodiments. The single Helix training stage coexists with prior internet and simulation pretraining. e-discussione-datae-architecturee-training |
6.3 Ablations and diagnostic examples
No ablation or diagnostic visual is included in this edition. The source-specific visual limitations are recorded above.
7. Analysis & limitations
7.1 What the evidence leaves open
Results are narrative accounts associated with videos. No quantitative outcome table, confidence interval, controlled ablation or systematic failure analysis is provided. The causal contributions of semantic pretraining, latent conditioning and latency-matched training therefore remain unresolved by this overview. e-upper-bodye-collaboratione-graspinge-traininge-designe-discussion
Excluding handled training objects supports an object-disjoint evaluation claim, but does not establish semantic categories absent from internet pretraining. Likewise, the source’s statement about no explicit role assignments needs qualification: its example instructions explicitly specify giving, receiving, left and right. Shared weights alone do not show spontaneous role allocation. e-datae-architecturee-collaboration
The reported control scope is the upper body. The overview supplies no locomotion evaluation, long-term household reliability study or measured power budget supporting its commercial-readiness claim. Its proposed future scaling by 1,000 times is an aspiration, not an evaluated configuration. e-introductione-upper-bodye-inferencee-conclusion
Figure 1 plots Number of tasks against Time with uncalibrated illustrative curves and no measured samples or uncertainty. Its claim about scaling behaviors without training concerns new behavior specification after learning; approximately 500 hours of robot supervision and component pretraining are explicitly reported. It cannot establish a scaling law, speedup factor or elimination of training cost. e-probleme-datae-architecturee-training
The state naming difference between Joint angles in the original diagram and wrist pose in the Architecture prose remains unresolved. Likewise, 20 Hz diagram inputs and the 200 Hz output loop are distinct rates. No source timing breakdown justifies assuming fresh sensory input for every output or inferring an interpolation mechanism. e-architecture-visuale-architecturee-inference
7.2 Questions for discussion
- How robust is semantic conditioning when the scene changes faster than System 2 updates?
- How much novel-object grasping transfers to new semantic categories, rather than new instances of familiar categories?
- Would shared weights still support collaboration without instructions that specify each robot’s role?
- What state transform and observation schedule reconcile the diagram's Joint angles and 20 Hz input labels with the prose's wrist poses and 200 Hz control loop?
8. Reproducibility audit
8.1 Requirements and known gaps
Reconstruction would require the specific VLM checkpoint, simulation-pretrained vision backbone, teleoperation data and hindsight-labeling procedure, robot action/state conventions, and asynchronous runtime. The overview does not identify checkpoints, optimizer, learning rate, training compute, exact latency offset, GPU models, actuator controller or termination rule. It also leaves the diagram/prose state mapping and handling between 20 Hz inputs and 200 Hz outputs unspecified. Mentioning an open-weight VLM does not establish availability of Helix weights. e-architecturee-actionse-datae-traininge-inferencee-architecture-visual
Proposed check, not run: train otherwise matched models with zero temporal offset versus an offset calibrated to measured deployment latency. Under a controlled S2-delay sweep, hold the sensory and action schedules fixed, log observation/latent age at every action, and measure grasp success as successful grasps divided by all attempted grasps on a predefined held-out object set. Compare response to moving targets separately, with repeated trials and uncertainty. A consistent advantage for the calibrated model would support the latency-matching rationale; no difference would weaken it. e-traininge-inferencee-architecture-visuale-data
Proposed check, not run: compare correct, shuffled and fixed S2 latents on identical held-out-object scenes with contrasting language commands. Keep S1 observations and timing unchanged. Report correct target selections divided by all commands, and successful grasps divided by all attempts, with repeated trials and uncertainty. Preserved motor competence with degraded instruction following under shuffled latents would support semantic conditioning; unchanged instruction following would weaken that interpretation. This is a reader proposal, not a reported ablation. e-architecturee-grasping
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: Does the training offset improve tolerance to semantic delay?
Proposed check, not run: train otherwise matched models with zero temporal offset versus an offset calibrated to measured deployment latency. Under a controlled S2-delay sweep, hold the sensory and action schedules fixed, log observation/latent age at every action, and measure grasp success as successful grasps divided by all attempted grasps on a predefined held-out object set. Compare response to moving targets separately, with repeated trials and uncertainty. A consistent advantage for the calibrated model would support the latency-matching rationale; no difference would weaken it. e-traininge-inferencee-architecture-visuale-data
Check 2: Does the latent carry instruction-specific information?
Proposed check, not run: compare correct, shuffled and fixed S2 latents on identical held-out-object scenes with contrasting language commands. Keep S1 observations and timing unchanged. Report correct target selections divided by all commands, and successful grasps divided by all attempts, with repeated trials and uncertainty. Preserved motor competence with degraded instruction following under shuffled latents would support semantic conditioning; unchanged instruction following would weaken that interpretation. This is a reader proposal, not a reported ablation. e-architecturee-grasping
8.3 Reading coverage
Visual audit: Read both complete primary text chunks using the supplied reader. Requested and visually inspected all ten supporting-section images and both original-figure excerpts. inspectedSections preserves exactly the descriptor's supporting-section inventory and order; the two original-figure records remain in visuals with their exact htmlSource locators. htmlEvidence identifies supporting-section context supplied alongside those original figures: e-problem additionally depends on figure1; e-dual and e-architecture-visual additionally depend on architecture; e-media additionally depends on both. Those original images and their adjacent text must accompany the supporting sections in review. Inspected both final copied PNG assets at original resolution, retaining all original labels, the complete Figure 1 caption, adjacent text and disclosures. The title/date and separate copyright were verified in identity. Cross-checked the architecture's arrow directions, S2-to-S1 latent, GPU/model labels, S1's 20 Hz sensory branches, 7–9 Hz S2 branches and semantic updates, 200 Hz output and Joint angles/Finger positions against model-data and training prose. The state mapping and detailed sampling schedule remain unspecified. The helper verified and delivered coordinator-prepared images; this reader did not rerender them or launch a browser. The architecture still uses outer frame 200 at 100 fps (2 s) and independent GIF frame 0 (0 ms), not synchronized historical playback. No complete animation or external video was viewed. Delivery, source hashes and this audit do not constitute independent scientific approval.
Original HTML supporting sections inspected: identity, model-data, training, introduction, design-advantages, results-body-control, results-coordination, results-object-generalization, discussion, conclusion. Appendix coverage: not present.
Original raster assets and animation geometry remain the work of the source’s authors. Animation-derived stills disclose their component times, background and wrapper; they do not establish full-animation or video inspection. The surrounding reading notes are our own.
Text reading scope & known omissions
- Complete primary HTML text, chunks 1 and 2, including title and February 20, 2025 article date
- Introducing Helix; New Scaling for Humanoid Robotics; original Figure 1 and caption
- Helix: A "System 1, System 2" VLA for Whole Upper Body Control; design advantages
- Model and Training Details: Data; Architecture; Training; Optimized Streaming Inference
- Results: Fine-grained VLA whole upper body control; Zero-shot multi-robot coordination; Emergent "Pick up anything"
- Discussion: Helix's training is highly efficient; A single set of weights
- Conclusion; Video 1–6 captions; remaining supplied site/contact text
- Original identity excerpt, including separately extracted website copyright
Outside the original text pass
- External figure assets are not downloaded; HTML alt text and mathematical text are retained where available.
- Both primary text chunks were read completely. This is the supplied official technical overview, not a conventional research paper, codebase review or later Helix release. No primary PDF, supplement or appendix was supplied.
- The original acquisition recorded: External figure assets are not downloaded; HTML alt text and mathematical text are retained where available. The pinned media upgrade now supplies the original Figure 1 raster and a derived architecture still with adjacent source text. Other external image assets remain unavailable and uninspected.
- External robot-demonstration Videos 1–6 were unavailable and were not watched. Their captions and associated claims were read. Only the embedded GIF frame-zero components visible in the derived architecture still were inspected; neither complete GIFs nor the complete architecture animation were viewed.
- The architecture image uses the disclosed component times: outer Lottie frame 200 at 100 fps (2 s), with each embedded GIF independently at frame 0 (0 ms), on the source-parent #0C0C0C matte. This is not synchronized historical playback. The coordinator prepared it offline using the pinned player in a trusted wrapper; no source HTML or iframe scripts were executed by this reader.
- The exact title and canonical URL match the catalog. The original identity section visibly dates the article February 20, 2025 and separately shows © Figure 2026. It displays no individual author byline. Metadata therefore credits Figure only as Source organization, based on the visible copyright/rightsholder notice, not as an author. The catalog's Figure AI author attribution is not verified as an author byline. The copyright year is not the publication date; affiliations are unverified.
- No numbered revision history is established. The source hash and acquisition time identify the reviewed snapshot; they do not prove pixel identity with the website on its original publication date. Images preserve original content with the disclosed offline layout wrapper and local Arial font.
- Code, Helix weights, datasets and external linked resources were not inspected; no experiments were reproduced. Existing report claims were rechecked against the complete primary text and the inspected source excerpts.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e-identityHTML fragment identity: article title (characters 5577–5680), date (5710–5810), separate copyright notice (53068–53115); no original anchor
The original heading reads Helix: A Vision-Language-Action Model for Generalist Humanoid Control and the article date is February 20, 2025. The separate website notice reads © Figure 2026. No individual byline appears in the identity section; the notice identifies Figure in a rightsholder role, not an author role.
Go to primary source ↓e-introductionHTML fragment introduction: Introducing Helix, opening capability list and Video 1 caption; characters 5896–7696 and 8037–8293
Figure claims upper-body control, two-robot operation, novel-object handling, one shared set of task weights and onboard embedded-GPU deployment. Priority and immediate commercial readiness are asserted without a comparative evaluation here.
Go to primary source ↓e-problemHTML fragment figure1: New Scaling for Humanoid Robotics, original Figure 1, caption and adjacent paragraphs; characters 8302–11883. Supporting context: introduction, capability list, characters 5896–7696
The prose motivates language-specified household skills by the cost of expert programming and demonstrations. The original raster shows Number of tasks versus Time and contrasts PhD hours, collected data and language-enabled new behaviors. Its curves have no numerical calibration or measured samples; the caption describes the proposed scaling comparison.
Go to primary source ↓e-dualHTML fragment architecture: Helix: A "System 1, System 2" VLA for Whole Upper Body Control, adjacent source text; characters 13425–14995. Supporting architecture/training context: model-data, 16815–19311, and training, 19311–21669
S2 is an internet-pretrained onboard VLM operating at 7–9 Hz; S1 translates latent semantics into continuous actions at 200 Hz. The prose describes two complementary systems trained end to end to communicate.
Go to primary source ↓e-dataHTML fragment model-data: Model and Training Details, Data paragraph; within characters 16815–19311
Approximately 500 hours of diverse multi-robot, multi-operator teleoperation are labeled with hindsight instructions by a VLM processing segmented onboard-camera clips. All items handled during training are excluded from evaluation. No detailed split inventory is supplied.
Go to primary source ↓e-architectureHTML fragment model-data: Architecture, first two paragraphs; within characters 16815–19311
S2 is a 7B-parameter open-source, open-weight VLM pretrained on internet-scale data. It processes monocular images, wrist pose/finger-position state and commands, distilling one continuous latent. S1 is an 80M-parameter cross-attention encoder-decoder with a simulation-pretrained multiscale convolutional vision backbone. The projected S2 latent is concatenated with visual features along the sequence dimension.
Go to primary source ↓e-actionsHTML fragment model-data: Architecture, final paragraph; within characters 16815–19311
S1 outputs desired wrist poses, finger flexion/abduction and torso/head orientation targets at 200 Hz. A synthetic percentage-task-completion action predicts termination for sequencing behaviors; no threshold or lower-level actuator-controller specification is supplied.
Go to primary source ↓e-trainingHTML fragment training: Training, first two paragraphs; within characters 19311–21669
Continuous actions use a standard regression loss; gradients propagate from S1 to S2 through the conditioning vector. One Helix training stage uses shared weights without per-task action heads or fine-tuning. Training offsets S1/S2 inputs to match the deployed inference-latency gap; no exact offset or loss formula is provided.
Go to primary source ↓e-inferenceHTML fragment training: Optimized Streaming Inference, three paragraphs; within characters 19311–21669
Each robot has two low-power embedded GPUs, with S2 and S1 on dedicated GPUs. Asynchronous S2 updates shared-memory intent from the latest observation and command; S1 uses the latest available observation and latent in its 200 Hz loop. The source links deployment to the training offset and claims the speed of its fastest single-task imitation policies, without hardware model, power or matched timing measurements.
Go to primary source ↓e-upper-bodyHTML fragment results-body-control: Results, Video 3 caption and Fine-grained VLA whole upper body control; characters 21669–21792, 22110–22259 and 22268–23294
The source reports a 35-DoF action space at 200 Hz and describes coordinated head tracking, torso reach and fingers, explaining that head/torso motion changes visibility and reach. It asserts novelty relative to prior VLAs without a matched baseline, trial statistics or outcome table.
Go to primary source ↓e-collaborationHTML fragments results-coordination (characters 23611–23729 and 23738–24843) and introduction (Video 1 caption, 8037–8293)
Two robots reportedly store unseen groceries with identical weights and no robot-specific training. The text also claims no explicit role assignments, yet its example commands specify handing to the robot on the right or receiving from the left and placing the item in a drawer. No trial denominator, success rate, uncertainty or matched baseline is given.
Go to primary source ↓e-graspingHTML fragment results-object-generalization: Emergent "Pick up anything" and Video 5 caption; characters 24843–24945, 33028–34002 and 34319–34495
The text reports systematic testing with thousands of unseen household items in clutter. A toy cactus is reportedly selected for a desert-item instruction, followed by closest-hand selection and grasping. Exact object/trial counts, a success denominator, uncertainty and quantitative comparisons are absent.
Go to primary source ↓e-discussionHTML fragment discussion: Helix's training is highly efficient; A single set of weights; Video 6 caption; characters 34504–36029 and 36345–36441
The source reports approximately 500 hours of supervision as <5% of previously collected VLA datasets, without naming the reference dataset or denominator. It compares that scale to single-task imitation data and claims no dependency on multi-robot-embodiment collection or multiple training stages. It reiterates one S2/S1 weight set for picking, containers, drawers, refrigerators, handovers and novel objects, contrasting prior systems' specialized fine-tuning or action heads.
Go to primary source ↓e-conclusionHTML fragment conclusion: Conclusion and future-scaling paragraph; characters 36450–37967
The conclusion reiterates claimed language-conditioned generalization and describes scaling Helix by 1,000 times and beyond as future ambition, not a tested configuration.
Go to primary source ↓e-mediaHTML fragments figure1 and architecture; Video 1–6 captions in introduction, design-advantages, results-body-control, results-coordination, results-object-generalization and discussion
The original Figure 1 presents conceptual skill scaling; the architecture graphic presents S2-to-S1 conditioning. Video captions identify grocery collaboration, reactive adjustments, upper-body control, coordination, conceptual object selection and picking up the helix. Captions alone provide no continuous-motion or task-success verification; the complete overview supplies no original quantitative table or controlled-ablation visual.
Go to primary source ↓e-designHTML fragment design-advantages: design-advantage list and Video 2 caption; characters 15317–15530 and 15539–16815
Figure claims specialist behavioral-cloning speed with zero-shot generalization, and contrasts continuous high-dimensional control with prior action tokenization for lower-dimensional grippers. It emphasizes standard component architectures and independent iteration without a unified observation/action representation. No controlled tokenization or speed/generalization ablation is supplied.
Go to primary source ↓e-architecture-visualHTML fragment architecture: original desktop animation-derived still (source iframe characters 13039–13211), adjacent section 13425–14995; compare model-data and training fragments
The inspected original labels show a 7B pretrained VLM on GPU 2, an 80M transformer on GPU 1, and an S2-to-S1 Latent Vector link at 7–9 Hz. Image and state branches into S1 are labeled 20 Hz; the branches into S2 are labeled 7–9 Hz; the whole-upper-body output is labeled 200 Hz. Diagram state is Joint angles and Finger positions, whereas the Architecture prose uses wrist pose and finger positions. The source does not specify a mapping or a complete input-to-output timing schedule.
Go to primary source ↓8.5 Primary sources
Helix: A Vision-Language-Action Model for Generalist Humanoid Control ↗
HTML · 2,206 extracted words
Source fingerprint
baeff5980b8b7343852513cb4a93b71b894fc8f2d8c0a363313389599bd5b943