Helix
Figure AI, United States · February 2025
Figure AI, Inc.
- Parameters
- 7.1 B
- Figure states 7B for System 2 and 80M for System 1, which sums to about 7.08B across the two networks.
- GPU memory
- 13 GB
- weights at bf16, computed. Fits a 24 GB card
- Inference latency
- not published
- per action step
- Weights
- closed
- no public checkpoint
What it is
Helix is Figure AI's in-house vision-language-action model for its humanoid, announced in February 2025. It splits the problem in two: a 7B VLM (S2) that looks at the scene and the instruction at 7 to 9 Hz and emits a latent vector, and an 80M visuomotor transformer (S1) that consumes that latent and drives a 35-DoF upper body at 200 Hz. Both run onboard on two embedded low-power GPUs, which is the specific engineering claim that sets it apart from published VLAs that need a server GPU. It is trained on roughly 500 hours of teleoperation with hindsight language labels. For anyone making a build-or-buy decision, the decisive fact is that Helix is closed end to end, has no public benchmark result, and only runs on Figure's own robots.
Architecture
- Backbone
- Two networks in one stack. S2 is a 7B-parameter open-source, open-weight VLM pretrained on internet-scale data, running at 7 to 9 Hz for scene and language understanding. Figure does not name which VLM it is. S2 emits a latent vector that conditions S1.
- Action head
- S1 is an 80M-parameter cross-attention encoder-decoder visuomotor transformer that consumes the S2 latent plus its own visual and proprioceptive input and outputs continuous actions over a 35-DoF upper-body action space at 200 Hz.
- Control rate
- 200 Hz
- Parameters
- Figure states 7B for System 2 and 80M for System 1, which sums to about 7.08B across the two networks. These are the only parameter figures Figure publishes. The identity of the 7B VLM used for S2 is not disclosed, so its pretraining corpus cannot be audited.
- Pretraining data
- S2 starts from an open-source VLM pretrained on internet-scale data. Helix itself was trained on roughly 500 hours of high-quality multi-robot, multi-operator teleoperated behaviour, which Figure describes as under 5 percent of the dataset size used in prior VLA work. Natural-language instructions were attached in hindsight by an auto-labelling VLM.
- Embodiments
- Figure 02 humanoid, 35-DoF upper body including individual fingers, wrists, torso and head, Two Figure 02 robots collaborating on a shared task using one identical set of weights
What hardware it needs
This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.
| Precision | Weights | Note |
|---|---|---|
| fp32 | 26 GB | Training master weights, rarely used for inference |
| bf16 or fp16 | 13 GB | The usual way these checkpoints are served |
| int8 | 6.6 GB | Quantised, expect some loss of precision on fine motions |
| int4 | 3.3 GB | Aggressive quantisation, verify success rate before trusting it |
Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 24 GB. That is an estimate for inference, not for fine tuning. Full fine tuning also holds optimiser state and gradients, which typically costs several times the weights unless you use LoRA or a comparable adapter.
| Control rate | 200 Hz |
Figure states Helix runs entirely onboard on "dual low-power-consumption embedded GPUs", with S2 and S1 deployed model-parallel across the two processors. Figure does not name the GPU model, its memory or any millisecond latency figure. The only rate figures published are 7 to 9 Hz for S2 and 200 Hz for S1.
Reported results
Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.
Simulation
No results in this category are published for this model.
Real world
Rollouts on physical hardware. The setups differ, so read these as evidence, not as a ranking.
- Figure internal logistics evaluation (parcel induction) Figure 02 fleet on a customer site, June 2025 update95%shipping labels correctly oriented for scanningVendor-reported, approximate figure, up from roughly 70 percent. No third-party verification.source
- Figure internal ablation (stereo vision) Logistics package handling, February 2025 update60%throughput increase of the stereo model over the non-stereo baselineVendor-reported relative improvement, not an absolute score. In the same post Figure reports 40 percent better throughput from curated demonstrations trained on one third less data, and states that 8 hours of well curated demonstration data can yield a workable policy.source
- Figure internal data-scaling study Same task, demonstration data scaled from 10 hours to 60 hours4.31 saverage processing time per package at 60 hours of dataVendor-reported. 6.84 s at 10 hours of demonstrations, which Figure describes as a 58 percent increase in throughput.source
- Figure internal logistics evaluation (parcel induction) Figure 02 fleet on a customer site, June 2025 update4.05 saverage cycle time per packageVendor-reported, no public benchmark and no third-party verification. Down from roughly 5.0 s in the previous update.source
Fine tuned tasks
No results in this category are published for this model.
On real hardware
Figure reports Helix running fully onboard Figure 02 humanoids. Demonstrated behaviours include picking up thousands of novel household items on natural-language request, two robots collaborating on a grocery put-away with one shared set of weights, and a production logistics deployment sorting parcels. In the June 2025 logistics update Figure reports 4.05 seconds per package (down from about 5.0 s) and about 95 percent correct label orientation (up from about 70 percent), with architectural additions of a vision memory module for stateful perception, state history and force sensing in the state input. All figures are vendor-reported and none are independently verified.
Fine tuning it yourself
Not possible for third parties. Figure releases no weights, no code and no API. Nothing about the training recipe beyond the 500-hour figure and the hindsight VLM auto-labelling is documented.
Where it helps, where it does not
Strengths
- Runs entirely onboard on embedded low-power GPUs with no cloud round trip, which is the concrete difference from most published VLAs that need a datacenter GPU in the loop.
- The split into a slow 7 to 9 Hz reasoner and a fast 200 Hz visuomotor policy decouples semantic generality from control rate, so a large VLM does not throttle the servo loop.
- Trained on roughly 500 hours of teleoperation, which Figure states is under 5 percent of the data used by comparable VLA efforts.
- One set of weights covers pick and place into containers, drawers and refrigerators, and two-robot handovers, without per-task heads.
Limits
- Fully closed. No weights, no code, no paper, no API and no reproducible evaluation. Every number on this row comes from Figure's own blog posts.
- No public benchmark result exists. Helix has never been scored on LIBERO, SimplerEnv, CALVIN or any other shared suite, so it cannot be ranked against open models on equal terms.
- Bound to Figure hardware. There is no way to fine-tune or deploy Helix on another robot.
- This first version controls the upper body only, 35 DoF. Locomotion is not part of it and was only added with Helix 02.
- Figure does not name the 7B VLM used for S2, so the model's licensing chain and pretraining data cannot be verified.
- No VRAM, no memory footprint and no millisecond latency figure is published, only the 7 to 9 Hz and 200 Hz loop rates.
Sources
Everything on this page was taken from these documents. Where they disagree with what you read here, they win.
- https://www.figure.ai/news/helix
- https://www.figure.ai/news/helix-logistics
- https://www.figure.ai/news/scaling-helix-logistics
- https://www.figure.ai/news
Entry last checked 2026-08-11.
Try a policy on a real arm
A physical SO-100 is online and free to drive in the browser, and five of the models in this arena can be fine tuned on a dataset you record with your own.