Helix 2.5 tests crowd video as a shortcut to home robot skills

Figure AI is sidestepping slow house-by-house robot demonstrations by converting crowdsourced domestic video into training data. According to geekpark, the company collects human behavioral data for its Index dataset at a rate of approximately 35 minutes of experience per second. This intake replaces physical hardware trials with consumer uploads. A web platform called Kled already structures this exchange, paying ordinary contributors for mundane first-person clips.

Data volume accumulates quickly under piece-rate bounties. Yet converting passive observation into safe domestic autonomy remains an unsolved physical hurdle.

Commercial platforms are proving that people will sell their private environments for modest sums. As woshipm reported, Kled listed task prices in the US market including $5 for filming trash disposal and $1,000 for uploading 500 selfies. During the first three months of 2025 as a web product, Kled paid users across five batches amounting to $52,000, $64,000, $75,000, $150,000, and $350,000, totaling nearly $700,000. Gathering human habits is now an economic reality, but translating cheap pixels into dependable home manipulation poses a starker operational test.

Crowdsourced domestic video builds a low-cost data pipeline

Figure AI quietly launched a data collection portal in May before unveiling it under the name Index on August 25, according to ifanr. At launch, the application had recorded 264,000 downloads across 108 countries alongside over 44,000 weekly active creators. Creators had supplied more than 16 million uploaded videos. According to geekpark, Figure AI gathers human behavioral data for Index at a rate of approximately 35 minutes of experience per second. The company has already paid $15 million to contributors, with plans to spend over $1 billion on data and compute over the next 12 months.

Individual payouts remain modest. In May, a reporter earned $21.55 uploading over 100 first-person daily life videos across Kled, Luel, and Waffle Video over one week, woshipm reported.

Compensation varies sharply by region and submission type. In the US market, Kled listed task payouts including $5 for filming trash disposal and $150 for uploading a tax filing document. Founder Avi Patel stated that one top-earning truck driver makes up to approximately $8,000 per month uploading dashcam footage and road pothole data. Ordinary users in Malaysia and the Philippines earn between $20 and $40 per month. By March, Kled had organized these submissions into more than 12,000 structured datasets.

Those datasets carry sweeping commercial permissions. Under terms of service updated on July 21, 2026, contributors retain ownership while granting Kled a perpetual and irrevocable license to train AI models. As woshipm author Zhang Aila argues, purchasing private domestic records introduces escalating compliance and privacy risks that consent alone cannot resolve.

Paid Crowdsourced Video Data Platforms for AI and Robotics

Transferring egocentric human clips to zero-shot robot actions

Figure AI released its Helix 2.5 model on September 17, attempting to replace manual, house-by-house robot teaching with passive human footage. Helix 2.5 begins with random initialization and pre-trains on Figure AI's human behavioral dataset Index before adapting to specific manipulation tasks. To test zero-shot transfer, engineers deployed the humanoid robot across 30 rented homes in the San Francisco Bay Area without prior environment-specific tuning or data collection. The setup removes human operators from the deployment loop entirely.

Figure AI CEO Brett Adcock described the 30-home deployment as the company's most important project to date. In those unmapped spaces, the robot completed 237 of 420 trial tasks, recording a 56% overall success rate.

That mark gains its significance when set against a control policy. According to geekpark and ifanr, a baseline model running identical hardware and task data without Index pre-training scored 9%. The video pre-training drove the difference even though no single evaluation task exceeded 1.90% of the Index dataset. Figure AI claimed a transfer scaling law after evaluating Index volumes at 1x, 2x, 4x, and 8x multiples. Results varied by chore. Bed-making reached a 67% success rate with 94 of 140 completions. Towel folding hit 62% on 87 of 140 runs, while toy tidying logged 40% on 56 of 140 attempts.

Not everyone views the results as proof of true adaptability. An author at ifanr expressed disappointment that the zero-shot evaluation included only three tasks. Nikolai Ensslen argued that the remaining 44% failure rate shows imitation learning systems are incapable of genuine generalization. For critics like Ensslen, visual mimicry leaves an autonomous agent stranded whenever real-world physics drift outside the distribution it observed.

The reliability gap between benchmark progress and household safety

In Figure AI's zero-shot test across 30 homes, Helix 2.5 completed 237 of 420 trial tasks, achieving an overall full-task success rate of 56%. Figure AI CEO Brett Adcock stated that the zero-shot home testing project is the company's most important project to date. According to geekpark, performance varied across the selected household chores: the robot achieved a 67% success rate (94 of 140) in bed-making, 62% (87 of 140) in towel folding, and 40% (56 of 140) in toy tidying.

A 44% failure rate makes autonomous deployment unworkable. Household chores offer little tolerance for dropped objects or stalled runs. Addressing the trial design directly, an author at ifanr expressed disappointment that Figure AI's zero-shot evaluation included only three tasks.

Industry peers argue that benchmark progress does not equal usable autonomy in lived environments. According to geekpark, Sunday Robotics co-founder and CEO Tony Zhao criticized Helix 2.5's 56% success rate, arguing that useful work requires both generalization and reliability. Specialized setups deliver far higher certainty. Sunday Robotics reported that its ACT-2 system completed 778 out of 785 autonomous single-garment folding trials, achieving a 99.1% reported success rate.

Nikolai Ensslen argued that Helix 2.5's 44% task failure rate serves as evidence that imitation learning systems are incapable of genuine generalization. That reliability gap becomes critical as hardware enters the commercial pipeline. Robotics company 1X opened pre-orders for its Neo home robot in February at a price of $20,000, setting a market expectation where buyers anticipate dependable daily execution rather than coin-flip trials.

Data Pipeline and Architecture Selection for Domestic Humanoid Manipulation

• Targeting zero-shot generalization across diverse, unfamiliar domestic environments without environment-specific fine-tuning Adopt large-scale human behavioral video pre-training. Incorporating Figure AI's Index pre-training improved zero-shot trial success from a 9% baseline to 56% across 30 rented San Francisco Bay Area homes, even when no single evaluation task represented more than 1.90% of the pre-training dataset.
• Deploying robots for single-task, high-reliability commercial or home chores that cannot tolerate failure rates above 10% Avoid relying purely on general imitation learning or broad behavioral pre-training models. While Figure AI's Helix 2.5 achieved an overall 56% success rate (67% in bed-making, 62% in towel folding, and 40% in toy tidying, failing 44% of tasks overall), Sunday Robotics' ACT-2 achieved a 99.1% success rate (778 of 785 trials) by targeting autonomous single-garment folding.
• Evaluating crowdsourced user-uploaded video and private data pipelines for model pre-training Weigh rapid volume expansion against acute privacy and licensing exposure. Platforms like Kled scaled to 200,000 daily active users, 3 million to 4.5 million daily files, and over 12,000 structured datasets by paying users for everyday activities (e.g., $5 for filming trash disposal), but commercializing private uploads-such as paying $150 for tax filings or $1,000 for 500 selfies-introduces mounting compliance, trust, and legal risks that standard consent terms cannot fully mitigate.
• Planning compute and infrastructure capacity for human-to-robot transfer scaling Budget compute aggressively alongside data collection when testing pre-training scaling laws (such as dataset volumes scaled across 1x, 2x, 4x, and 8x multiples). Sustaining an intake of approximately 35 minutes of behavioral experience per second requires massive compute backing, exemplified by Figure AI's multi-year Nscale partnership committing $3.5 billion to over $6 billion for up to 100,000 Nvidia Vera Rubin GPUs.

Capital requirements, compute bottlenecks, and capture protocols

Figure AI reached a post-money valuation of $39 billion after raising more than $1 billion in its latest funding round, according to geekpark. The company claims to have manufactured over 350 units of the Figure 03 humanoid robot and reduced its manufacturing cycle to one unit per hour. Physical assembly is not the blocker. Figure AI CEO Brett Adcock stated that data and compute are the company's biggest bottlenecks in deploying humanoid robots to homes worldwide, as reported by ifanr.

To ease compute constraints, Nscale agreed to supply an initial $3.5 billion in capacity, expanding to over $6 billion. Up to 100,000 Nvidia Vera Rubin chips will deploy in Barstow, Texas, in the second half of 2027.

Gathering raw domestic activity demands equal financial momentum. According to ifanr, Figure AI has paid $15 million to data creators and plans to spend over $1 billion on data and compute over the next 12 months. Capturing living rooms at scale, however, creates friction over ownership. Terms of service from Kled updated on July 21, 2026, stipulate that while users retain ownership of uploaded clips, they grant an exclusive, perpetual, transferable, and sublicensable license for model training and deployment.

Paying camera owners does not erase legal vulnerability. As woshipm author Zhang Aila argues, platforms offering high payouts for private personal data confront escalating compliance and privacy risks that user consent alone cannot resolve. Contracts fail to anticipate every bystander. When unscripted living spaces enter training corpuses, standard release agreements struggle to insulate autonomous robotics systems from statutory privacy disputes.

Figure AI paid $15 million to amass over 16 million videos through Index, yet ifanr's author noted that the public zero-shot evaluation tested only three tasks. Sourcing crowd footage creates volume quickly, but physical AI teams must structure task protocols before buying scale.

Compute constraints compound the challenge. Figure CEO Brett Adcock called compute and data the primary barriers to deploying domestic humanoids, driving plans to spend over $1 billion on them over the next 12 months.

That physical infrastructure takes years to build. Nscale's contract to deploy up to 100,000 Nvidia Vera Rubin chips in Barstow, Texas targets the second half of 2027, while competitors like 1X sell $20,000 Neo pre-orders today. Teams entering this space should track task diversity, not video counts.

For readers outside China

• Availability: Figure AI's crowdsourced data portal, Index, is available internationally, registering 264,000 downloads across 108 countries with over 44,000 weekly active creators as of its August 25 official launch. User-compensation platforms like Kled operate across multiple countries, including the United States, Malaysia, and the Philippines, surpassing 500,000 registered users in July. On the physical hardware side, commercial availability remains limited: Figure AI has manufactured over 350 units of the Figure 03 robot at a cycle of one unit per hour and tested its Helix 2.5 model across 30 San Francisco Bay Area homes, but wide consumer delivery dates are not disclosed in sources. Competing manufacturer 1X opened pre-orders for its Neo home robot in February.
• Pricing: Robot hardware pricing is currently anchored by 1X's Neo home robot, which opened pre-orders at $20,000; commercial unit pricing for Figure AI's hardware is not disclosed in sources. For crowdsourced data collection, Figure AI has paid $15 million to creators through Index and plans to allocate over $1 billion across data and compute over the next 12 months. On Kled, US task rates include $5 for filming trash disposal, $150 for submitting a tax filing document, and $1,000 for uploading 500 selfies. Monthly earnings vary heavily by location and role: ordinary users in the Philippines and Malaysia earn between $20 and $40 per month, a WIRED journalist earned $21.55 for uploading over 100 videos across Kled, Luel, and Waffle Video in one week, and high-volume contributors such as truck drivers uploading dashcam data earn up to approximately $8,000 per month. Backend compute spending is substantial, reflected in Figure AI's initial $3.5 billion compute contract with Nscale, which includes options expanding past $6 billion.
• Closest Western equivalents: Sunday Robotics (developer of the ACT-2 autonomous manipulation system); 1X (manufacturer of the Neo domestic humanoid robot); Figure AI (developer of Helix 02, Helix 2.5, Figure 03, and the Index platform); Nscale (GPU cloud infrastructure provider deploying Nvidia Vera Rubin hardware)
• Data residency: Terms of service for crowdsourced data platforms impose sweeping cross-border usage rights. Under Kled's terms updated on July 21, 2026, uploaders retain underlying ownership but must grant the company an exclusive, worldwide, perpetual, irrevocable, transferable, and sublicensable license to utilize content for AI model development, testing, and deployment. On the infrastructure side, Figure AI's large-scale compute operations with Nscale are anchored in Barstow, Texas, with deployment scheduled to begin in the second half of 2027. Dedicated local data-residency guarantees, regional storage boundaries, and GDPR compliance architectures for international uploads originating from the 108 active countries are not disclosed in sources.

Sources

• woshipm 50万人把自己的照片和视频卖给 AI，机器人训练也有"零工经济"了 https://woshipm.com/it/6453441.html
• ifanr Figure 让机器人第一次「空手」进陌生人家：不遥操，也不提前预习，进了家门就干活 https://ifanr.com/1680884
• geekpark Figure AI 宣称找到了机器人版 scaling law，同行却说它其实根本不会泛化 https://geekpark.net/news/370607