9c8a8ba4eb
Some checks are pending
Mirror PR to Forgejo / mirror (pull_request) Waiting to run
Pipeline auto-fixer: removed [[ ]] brackets from links that don't resolve to existing claims in the knowledge base.
81 lines
7.1 KiB
Markdown
81 lines
7.1 KiB
Markdown
---
|
|
type: source
|
|
title: "Figure 03 + Helix 02: Kitchen Demo and Manipulation Breakthrough in Unstructured Environments (January-February 2026)"
|
|
author: "Figure AI (@FigureAI)"
|
|
url: https://www.figure.ai/news/helix-02
|
|
date: 2026-01-28
|
|
domain: robotics
|
|
secondary_domains: [manufacturing]
|
|
format: thread
|
|
status: unprocessed
|
|
priority: high
|
|
tags: [humanoid-robots, Figure-03, Helix-02, manipulation, unstructured-environments, kitchen-demo, tactile-sensing, full-body-autonomy]
|
|
intake_tier: research-task
|
|
---
|
|
|
|
## Content
|
|
|
|
**Helix 02 overview (released January 28, 2026):**
|
|
|
|
Figure AI unveiled Helix 02, a full-body visuomotor neural network that replaces all prior C++ control architecture with a unified AI system. Architecture breakdown:
|
|
- S0: 10M-parameter neural prior at 1 kHz (low-level motor control)
|
|
- S1: Unified visuomotor at 200 Hz (whole-body coordination)
|
|
- S2: Semantic reasoning layer (task understanding)
|
|
- ALL C++ code from Helix 01 BMW deployment eliminated — the architectural ceiling identified at BMW is resolved
|
|
|
|
**Kitchen demo (4-minute autonomous task, 61 loco-manipulation actions):**
|
|
- End-to-end autonomous execution, no human resets, no hard-coded sequences
|
|
- Walk to dishwasher → unload dishes → navigate across kitchen → stack in cabinets → reload dishwasher → start cycle
|
|
- Tests: walking, grasping, object recognition, spatial navigation, sequenced multi-step planning, all integrated
|
|
|
|
**Dexterous manipulation capabilities demonstrated:**
|
|
- Tactile fingertip sensors: 3-gram force detection ("sensitive enough to feel a paperclip")
|
|
- Pill extraction from medicine organizer (uses palm-level visual feedback)
|
|
- 5mL syringe actuation (force-controlled to exact volume)
|
|
- Cluttered box singulation (objects overlapping, shifting during interaction, self-occlusion)
|
|
- Unscrewing bottle cap (bimanual coordination with tactile-regulated grip force)
|
|
|
|
**Hardware improvements (Figure 03 vs. Figure 02):**
|
|
- Camera: 2x frame rate, 1/4 latency, 60% wider field of view per camera
|
|
- More compact form factor
|
|
- Embedded tactile sensing in each fingertip and palm cameras (new hardware)
|
|
|
|
**Figure 03 and BotQ factory:**
|
|
- BotQ facility (Sunnyvale, California): dedicated high-volume humanoid manufacturing
|
|
- Production ramp: 1/day → 1/hour (24x improvement in under 120 days, announced ~May 2026)
|
|
- Units delivered: 350+ Figure 03 robots to partners/pilots by May 2026
|
|
- Current pace: ~55 robots/week
|
|
- First-pass yield: 80% at BotQ
|
|
- Infrastructure: 150 networked workstations with custom MES
|
|
- Target capacity: 12,000 units/year initially; 100,000 total over 4 years
|
|
- Consumer pricing target: $20,000 (aggressive; requires significant manufacturing scale)
|
|
- Home deployment timeline: select partner testing in 2026; broader consumer availability late 2026 to 2027
|
|
|
|
**From Time Magazine (Figure 03 profile):** "Figure 03 Is The Robot in Your Kitchen" — framing the consumer market as the target
|
|
|
|
**From GoPenAI/Medium (May 2026 analysis):** "Figure Just Solved the Hardest Problem in Robotics" — referring to unstructured manipulation
|
|
|
|
## Agent Notes
|
|
|
|
**Why this matters:** This is the leading indicator I flagged in the May 11 musing for the manipulation constraint crossing. The May 11 session identified three binding constraints on humanoid robot deployment: hardware reliability, software architecture generalization, and manipulation in unstructured environments. Helix 02 / Figure 03 addresses all three: the C++ architectural ceiling is resolved (architecture), the kitchen demo demonstrates genuine unstructured manipulation (capability), and BotQ's 80% first-pass yield suggests manufacturing maturity (reliability). This is the most significant robotics development of May 2026.
|
|
|
|
**What surprised me:** The specific manipulation tasks — pill extraction, syringe force control, cluttered box singulation — are not structured factory tasks. These are healthcare and household ADL tasks. Figure is targeting the home market directly, not the factory market. This is a different commercial thesis than Figure 02's BMW deployment. It also means the "kitchen is still more structured than full unstructured" objection is weakening — healthcare manipulation in particular is high-variability.
|
|
|
|
**What I expected but didn't find:** Expected the consumer deployment timeline to be 2027+. The BotQ 24x production ramp and 350+ units delivered by May 2026 suggests the scale-up is proceeding faster than I anticipated. The $20K price target is still aspirational — current units are being deployed to partners, not sold at consumer prices. The cost threshold crossing is still 2027+ at earliest.
|
|
|
|
**KB connections:**
|
|
- Belief 11 (robotics is binding constraint): The manipulation constraint — the hardest of the three — is being meaningfully breached. The "unsolved" characterization from prior sessions needs updating.
|
|
- three conditions gate AI takeover risk autonomy robotics and production chain control and current AI satisfies none of them — the robotics condition is now further along toward crossing than the prior framing
|
|
- knowledge embodiment lag means technology is available decades before organizations learn to use it optimally — Figure's C++ → full-body neural network transition is knowledge embodiment lag in reverse: the knowledge IS being embodied, rapidly
|
|
|
|
**Extraction hints:**
|
|
- CLAIM CANDIDATE 1: "Figure AI's Helix 02 demonstrates autonomous kitchen-task execution across 61 loco-manipulation actions including pill extraction, force-controlled syringe operation, and cluttered-object singulation — the first credible evidence that unstructured domestic manipulation is achievable by humanoid robots"
|
|
- CLAIM CANDIDATE 2: "Figure 03's BotQ manufacturing facility achieved 24x production throughput improvement (1/day to 1/hour) with 80% first-pass yield in under 120 days, demonstrating that humanoid robot manufacturing is entering serial production at commercially relevant volumes"
|
|
- SCOPE NOTE: Capability breakthrough ≠ cost threshold crossing. $20K consumer price target requires further manufacturing scale. The manipulation capability is demonstrated; the economics at mass-market scale are not yet closed.
|
|
|
|
**Context:** Figure AI is a California-based humanoid robotics company founded by Brett Adcock. Valuation as of early 2026: $39B. BMW deployment (Figure 02, 30,000 vehicles, 1,250 hours) was Gate 1b commercial validation (see prior archive). Helix 02 is the direct successor, released after BMW deployment lessons.
|
|
|
|
## Curator Notes (structured handoff for extractor)
|
|
PRIMARY CONNECTION: Belief 11 (robotics is binding constraint) — specifically the manipulation-in-unstructured-environments constraint identified as the hardest gate
|
|
WHY ARCHIVED: First credible public demonstration of domestic unstructured manipulation capability by a humanoid robot. The kitchen demo (61 actions, 4 minutes, autonomous, no resets) is materially more complex than prior humanoid demonstrations. Combined with BotQ production ramp, this is the leading indicator for the manipulation constraint crossing.
|
|
EXTRACTION HINT: Two claims, not one: (1) capability demonstration (what Helix 02 can do), (2) manufacturing ramp (what BotQ is achieving). Keep separate — they address different constraints.
|