Introducing Helix 02: Full-Body Autonomy
Source
- Raw Markdown: paper_helix-02-2026.md
- Official technical writeup: Introducing Helix 02: Full-Body Autonomy
- Prior source: Helix
Core Claim
Helix 02 extends Figure’s Helix hierarchy from upper-body control to full-body humanoid loco-manipulation. The official writeup describes a three-system stack from semantic context to visuomotor joint targets to high-rate balance/contact execution.
Method Notes
- System 2 handles scene understanding, language, goals, and behavior sequencing.
- System 1 remains a Transformer conditioned on System 2 latents and outputs full-body joint targets at 200 Hz.
- System 0 is a learned whole-body controller, described as a 10M-parameter neural network that outputs joint-level actuator commands at 1 kHz.
- Observations include head cameras, palm cameras, fingertip tactile sensors, and full-body proprioception; outputs cover legs, torso, head, arms, wrists, and fingers.
- The official source does not state that Helix 02 uses diffusion, flow matching, or a regression loss, so those labels should not be inferred.
Evidence And Limitations
Figure reports a 4-minute autonomous dishwasher task, 61 ordered loco-manipulation actions, bimanual transfers, dexterous contact tasks, and all videos as autonomous rather than teleoperated. The evidence is still company-published demonstration evidence with no public weights, dataset, ablations, benchmark protocol, or failure-rate statistics.
Foundation TSFM Relevance
| Agenda slot | Verdict | Evidence | Missing pieces |
|---|---|---|---|
| Dynamic compute allocation | adjacent | The S2/S1/S0 stack separates semantic planning, trajectory-level control, and high-rate actuator control. | Company writeup only; no open model, dataset, ablations, or digital-system equivalent. |
| Streaming state and constant updates | adjacent | System 1 is described as producing joint targets at 200 Hz and System 0 actuator commands at 1 kHz from continuous sensor inputs. | No explicit latent-state maintenance interface or long-horizon memory evaluation. |
| Benchmarks | warning | Autonomous demonstration evidence shows a plausible hierarchy but lacks failure-rate statistics. | Needs reproducible benchmark protocols and raw trajectory data. |
Links Into The Wiki
- Helix 02
- Helix
- Foundation Time-Series Model Research Agenda
- Robotics Time-Series Modeling
- Robotics Text Conditioning
Open Questions
- Should the S0/S1/S2 split be the canonical fast/slow abstraction for whole-body humanoids?
- How should this wiki compare company demonstration evidence against open paper benchmarks?