SenseTime open-sources SenseNova-RoboRSI framework and 27B embodied model

Recently, the SenseNova joint research team at SenseTime released a technical report on arXiv that systematically evaluated the real capabilities of OpenAI's GPT-6-Astra on embodied intelligence tasks, covering 12 simulation benchmarks and 8 real robot platforms. The test results themselves are already worth discussing: after relaxing the interaction step limit (increasing it to 10 times the standard budget) and allowing In-Context Learning, GPT-6-Astra approached or even reached the saturation line on multiple simulation benchmarks, with a task success rate of 81.9% on RoboDojo—a figure that corresponds to performance under a 10x step budget, while under the standard budget the success rate on the same benchmark was 27.6%. But on real robot platforms, long-horizon tasks, cross-task generalization, and fine control of deformable objects remain clear weaknesses.

What is more worth unpacking is another move SenseTime made at the same time: open-sourcing the SenseNova-RoboRSI execution framework and the accompanying 27B embodied model, claiming to reduce task completion costs to 1/50 of directly invoking GPT-6-Astra. According to SenseTime's disclosure, this figure comes from its own testing and awaits third-party replication; it measures the cost compression of specific tasks such as stacking bowls and loading pens under extended budget conditions, and is not a universal cost benchmark.

The cost of GPT-6-Astra in the extended-budget tests is quantifiable: stacking bowls took an average of 131 minutes, with inference costs of about $19; loading pens into a pen holder took an average of 263 minutes, with inference costs of about $55. 1/50 means that, for the same tasks, SenseTime's open-source solution can compress the cost to less than $1. But the scope of this conversion needs to be made clear: it measures cost compression for specific tasks under the extended budget conditions set by SenseTime, and is not equivalent to a universal cost benchmark across scenarios and task types.

This is not a purely academic open-source move. It comes two months after SenseTime just delivered its first IFRS profitable interim report since listing, and at a moment when the capital market remains sharply divided over the logic of its valuation. SenseTime is using an extremely pragmatic approach to answer a question it has long avoided: given that its core business has not yet fully achieved operating profitability, how can it rediscover its position within a larger technological narrative?

A strategic shift from "selling models" to "building an ecosystem"

SenseTime's financial structure in the first half of 2026 is the premise for understanding this move.

First-half revenue was 2.91 billion yuan, up 23.4% year on year; IFRS net profit was 617 million yuan, positive for the first time since listing. But the composition of this profit needs to be explained upfront: the main factor driving IFRS profitability was not a full turnaround in operating profit, but "net other gains" of 2.127 billion yuan, including a 521 million yuan gain from the disposal of subsidiaries and 1.249 billion yuan in unrealized gains on financial assets measured at fair value. Excluding these items, the Non-IFRS adjusted loss was still 386 million yuan, and the core business has not yet crossed the operating break-even line. In other words, the signal value of "first profitability" outweighs any substantive operational transformation.

At the same time, SenseTime's revenue structure has completed the switch from computer vision AI to generative AI: generative AI revenue was 2.327 billion yuan, accounting for 79.9% of the group's total revenue; traditional computer vision AI revenue was 497 million yuan, with its share shrinking to about 17%. Token service volume is expanding rapidly, and the large-scale computing infrastructure business is advancing the large-scale commercialization of domestic heterogeneous computing power, but the asset-heavy logic of the Token factory means that a gap between scale growth and profit improvement always exists and requires continuous verification.

At this juncture, embodied intelligence has been pushed to the forefront.

The logic is not complicated. The competitive landscape in generative AI is already highly concentrated, and SenseTime faces comprehensive pressure from rivals such as Alibaba, ByteDance, and DeepSeek in terms of model capabilities and pricing power. The Token factory business is essentially a hybrid of computing power leasing and model services; scale can be expanded, but margin flexibility is constrained by upstream computing costs and downstream customer bargaining power. Embodied intelligence—or "AI in the physical world"—remains a track where standards have not yet solidified and the landscape has not yet converged. Entering at this stage and using an open-source approach to build influence over the developer ecosystem and technical standards is a means of leveraging a long-term ecosystem position at extremely low cost. Strategically, this is an attempt to trade open source for an ecosystem position, and success depends on whether the community truly adopts it.

SenseTime's chosen entry point is not to build robot bodies, nor to compete in general-purpose embodied large-model benchmark rankings, but to "make embodied tasks cheap." The design philosophy of the RoboRSI framework is clear: package each simulation benchmark and its dependencies into an independent Docker image, expose environment interaction and native scoring interfaces through a unified HTTP API, and allow researchers to run end-to-end evaluation with just one configuration file, without manually setting up each simulator. The positioning of the 27B model is to "internalize" validated experience into model weights and reduce the number of trial-and-error attempts. According to SenseTime's disclosure, RoboRSI-27B improved the success rate on RoboCasa365 from 26.0% to 46.5% while reducing token usage by 73.3%; this result comes from SenseTime's own evaluation, and RoboCasa365 is a simulation benchmark, so its ability to transfer to real physical environments still needs verification.

The goal of this combination is not "the strongest" but "the most usable." As embodied intelligence moves from the lab to real-world scenarios, cost, rather than benchmark scores, is becoming the real bottleneck. Whether SenseTime's solution can truly address this bottleneck still depends on actual deployment and independent re-testing.

The substance of 1/50: the qualifying conditions behind the number

1/50 is a highly shareable number, but it needs to be understood under specific test conditions.

SenseTime's comparison baseline is the inference cost of GPT-6-Astra under "extended budget" conditions. A considerable portion of GPT-6-Astra's failure modes are not that it "cannot do it," but that it "stops too early"—among the 9 tasks in RoboDojo, it originally failed all 5 times, and merely by allowing the model to continue interacting and raising the step limit by 10x, it became 5 out of 5 successes. This means that a large part of GPT-6-Astra's inference cost is consumed by repeated trial and error and exploration with extremely long step counts.

The cost-reduction logic of RoboRSI is essentially replacing "trial and error at inference time" with "experience from pretraining." The 27B model solidifies task patterns that have been validated through evaluation into its weights, so that when facing new tasks the model does not need to start from scratch and interact repeatedly. SenseTime's own phrasing is to "build tested experience into the model weights so that it requires less trial and error."

This explains the substance of 1/50: it measures the cost compression of a specific class of embodied tasks (short-horizon manipulation such as stacking bowls and loading pens) under specific evaluation conditions (extended budget, simulation benchmark), rather than a universal, cross-scenario cost baseline. SenseTime also acknowledges that cost "is still a problem." In long-horizon tasks, cross-task transfer, and fine control on real robot platforms, the bottleneck has not disappeared just because a 27B model has been open-sourced.

Open source as a strategic tool: what SenseTime is betting on

SenseTime's choice to open-source at this point reflects specific strategic considerations.

First, the ecological niche of embodied intelligence has not yet been locked in. Unlike generative AI—which already has clear leading models and pricing systems—the software stack, evaluation standards, and execution frameworks of embodied intelligence are still at an early stage. Open-sourcing a reproducible evaluation framework and accompanying model at this stage is an effort to become "the default tool." Once RoboRSI becomes a common entry point for researchers re-testing and comparing embodied capabilities, SenseTime will have gained a non-exclusive form of influence in this track. The key lies in whether the community truly adopts it.

Second, open source is a low-cost means of hedging valuation pressure on the core business. SenseTime's market capitalization has shrunk by more than 80% from its peak after listing, and the market's patience with whether its "Token factory can be independently profitable" is still being consumed. Embodied intelligence provides a new narrative anchor: a company that accumulated perception and spatial understanding capabilities in the visual AI era has a kind of technological adjacency in the early stage of physical AI. Whether this adjacency can be converted into commercial competitiveness still requires commercial validation, but it at least provides the capital market with a differentiated label distinct from "yet another loss-making large-model company."

Third, SenseTime's financial structure determines that it cannot repeat the Token-factory-style heavy-asset investment in embodied intelligence. The R&D cost of an open-source framework and a 27B model is far lower than building its own ten-thousand-GPU cluster to train a hundred-billion-parameter embodied large model. This is a playbook that substitutes engineering efficiency and ecosystem strategy for capital density, and it matches SenseTime's current resource constraints.

Conclusion: from "proving it can make money" to "proving it still has a place"

SenseTime's 2026 story is undergoing a shift in focus.

The IFRS profitability in the first half of the year addressed the question of "whether this company will collapse financially"—more precisely, "whether it will collapse at the reporting level," rather than "whether the core business already has the ability to generate its own cash flow." The Non-IFRS loss narrowed to 386 million yuan, Token service volume is expanding, and the large-scale facility is advancing the large-scale commercialization of domestic computing power—these are progress at the survival level. But pricing in the capital market has never been only about survival. It is about whether a company still holds an irreplaceable position in the next stage of technological evolution.

The open-sourcing of RoboRSI is SenseTime's first formal answer to this question. The substance of that answer does not depend on how much impact the number 1/50 has in dissemination, but on three things: whether this framework can be truly adopted by the embodied intelligence research community, whether the generalization performance of the 27B model in actual robot tasks can withstand independent re-testing, and whether SenseTime can find a path between the open-source ecosystem and commercial monetization that does not rely on the expansion of Token volume.

The window for embodied intelligence will not stay open forever. The time SenseTime has bought with open source needs to be cashed in through the depth of its ecosystem. And before that happens, 1/50 remains a promise that needs to be verified, not a conclusion that has already been delivered.