Figure Claims a Robotics Scaling Law

Read original on 极客公园

#humanoid-robots#generalization#imitation-learningfigure-helix-2.5figure aihelix 2.5sunday roboticsindex

Figure’s 30-home test challenges whether robot learning truly generalizes beyond demos.

30-Second TL;DR

What Changed

Helix 2.5 performed household tasks in 30 homes without home-specific robot data or fine-tuning.

Why It Matters

The test is a meaningful attempt to measure household generalization outside controlled demos, but a 56% success rate is not yet sufficient for dependable domestic deployment. The debate also shows that robotics scaling claims need standardized, independently verified evaluations.

What To Do Next

Reproduce the evaluation design with held-out homes, objects, and lighting conditions, and report per-task failures rather than only aggregate success.

Who should care:Researchers & Academics

Key Points

•Helix 2.5 performed household tasks in 30 homes without home-specific robot data or fine-tuning.
•The evaluation covered tidying toys, folding towels, and making beds.
•The reported 56% task success rate triggered criticism from robotics competitors.

Key numbers$15 million56%9%50%

Deep Insight

Background and context from public sources — not the original article. 14 sources cited.

Enhanced Key Takeaways

•Figure AI's Helix 2.5 relies on its crowdsourced 'Index' data ecosystem, which has amassed over 16 million domestic activity videos by paying contributors more than $15 million and ingesting 35 minutes of footage per second.
•The 56% zero-shot home completion rate represents a 6x performance increase compared to models trained without human video pre-training, which achieved only a 9% success rate.
•Helix 2.5 cut task-specific fine-tuning data requirements by 50% relative to Helix 02, while no single test task accounted for more than 1.90% of the pre-training distribution.
•To power its transfer scaling law, Figure entered a $3.5 billion infrastructure agreement with Nscale to secure access to up to 100,000 Nvidia GPUs along the Vera Rubin roadmap.
•Robotics competitors like Sunday Robotics co-founder Tony Zhao criticized the 44% failure rate as unviable for fragile residential settings, noting their specialized ACT-2 architecture achieves 99.1% success on garment manipulation.

Competitor Analysis

Figure AI (Helix 2.5)

Primary Training Paradigm: Human-to-humanoid video pre-training (Index dataset) + sparse adaptation
Benchmark / Deployment Record: 56% complete-task success across 30 unmapped homes (237/420 trials)
Generalization vs. Reliability Trade-off: High zero-shot environmental breadth, but high (44%) error rate in unstructured consumer settings

Sunday Robotics (ACT-2)

Primary Training Paradigm: Targeted imitation learning / Action Chunking Transformer
Benchmark / Deployment Record: 99.1% success on domestic garment manipulation tasks
Generalization vs. Reliability Trade-off: High reliability on constrained manipulation routines, but narrower cross-task zero-shot generalization

Company / Model	Primary Training Paradigm	Benchmark / Deployment Record	Generalization vs. Reliability Trade-off
Figure AI (Helix 2.5)	Human-to-humanoid video pre-training (Index dataset) + sparse adaptation	56% complete-task success across 30 unmapped homes (237/420 trials)	High zero-shot environmental breadth, but high (44%) error rate in unstructured consumer settings
Sunday Robotics (ACT-2)	Targeted imitation learning / Action Chunking Transformer	99.1% success on domestic garment manipulation tasks	High reliability on constrained manipulation routines, but narrower cross-task zero-shot generalization

Technical Deep Dive

Data Ingestion Engine (Index): Helix 2.5 leverages crowdsourced egocentric and third-person domestic video streams capturing unconstrained real-world behavior, scaling ingestion throughput to approximately 35 minutes of human activity per second.
Human-to-Humanoid Transfer Scaling Law: The architecture relies on an empirical power-law relationship where action prediction loss scales down predictably with an 8-fold increase in pre-training data, while keeping downstream adaptation capacity fixed.
Pre-Training Distribution Guardrails: Zero-shot generalization was validated by restricting any single evaluation task (e.g., bed making, towel folding) to under 1.90% of the overall pre-training corpus distribution.
Compute & Training Infrastructure: Backed by an Nscale multi-year cloud contract provisioned for up to 100,000 Nvidia GPUs (targeting the Vera Rubin generation) to handle continuous multi-modal video tokenization and embodied policy convergence.

Future ImplicationsAI analysis grounded in cited sources

Human video pre-training will surpass teleoperation as the dominant scaling vector for humanoid foundation models.

Empirical proof of action prediction loss decreasing with doubled human demonstration data makes passive video ingestion dramatically more scalable and cost-effective than physical teleoperation rigs.

Humanoid residential deployments will remain commercially blocked without closed-loop error recovery mechanisms.

A 44% failure rate poses severe safety and liability risks in consumer environments, necessitating hybrid architectures that marry scaling laws with deterministic recovery policies.

Timeline

2024-02

Figure AI closes $675M Series B funding at a $2.6B valuation to accelerate embodied AI development

2024-08

Figure unveils its second-generation humanoid robot hardware platform, Figure 02

2026-08

Figure launches the Index crowdsourced video collection ecosystem out of stealth

2026-09

Figure secures $3.5B compute partnership with Nscale for up to 100,000 Nvidia GPUs

2026-09

Figure reveals Helix 2.5 and asserts the human-to-humanoid robotics scaling law across 30 test homes

2024-02
Figure AI closes $675M Series B funding at a $2.6B valuation to accelerate embodied AI development
2024-08
Figure unveils its second-generation humanoid robot hardware platform, Figure 02
2026-08
Figure launches the Index crowdsourced video collection ecosystem out of stealth
2026-09
Figure secures $3.5B compute partnership with Nscale for up to 100,000 Nvidia GPUs
2026-09
Figure reveals Helix 2.5 and asserts the human-to-humanoid robotics scaling law across 30 test homes