General Trajectory: Teaching AI the Physical World

General Trajectory is an early-stage artificial intelligence startup founded in 2024 with an unusually ambitious scope: bringing AI out of purely digital environments and into the physical world. Based in San Francisco and part of the Winter 2025 startup batch, the company is building foundational technology aimed at enabling machines—specifically humanoid robots—to perform economically meaningful work in real-world environments.

While most AI systems today are optimized for text, images, or abstract decision-making, General Trajectory operates on the belief that true general intelligence must interact with the physical world. Manufacturing, logistics, healthcare, defense, construction, and scientific research all depend on physical labor and manipulation. According to the company’s worldview, these sectors represent roughly 80% of global GDP—yet remain largely untouched by modern AI capabilities.

General Trajectory exists to close that gap. Its mission is not incremental automation, but a deeper rethinking of how machines learn to act, adapt, and generalize in unpredictable real-world environments.

Who Is Behind General Trajectory and What Shaped Its Direction?

The company was founded by Joshua Belofsky, a recent graduate with a background in academic machine learning research. Unlike many founders who arrive from industry or serial entrepreneurship, Belofsky comes from a research-driven perspective, with a focus on foundational problems rather than product polish.

This academic grounding strongly influences General Trajectory’s approach. Rather than targeting narrow robotic tasks or single-use automation, the company focuses on building a foundation model—a generalizable learning system capable of transferring skills across tasks, objects, and environments.

As a solo founder at launch, Belofsky represents a growing class of AI researchers turning directly to startup formation as a way to accelerate progress. The presence of Harj Taggar as a primary partner adds institutional credibility and strategic support, helping bridge the gap between deep research and real-world deployment.

What Problem in Robotics Is General Trajectory Actually Solving?

Despite dramatic advances in robotics, a fundamental paradox remains: robots can perform visually impressive feats like backflips, parkour, or choreographed movements, yet fail at simple, economically useful tasks such as picking up unfamiliar objects or manipulating tools reliably.

Dexterous manipulation—the ability to grasp, move, and interact with objects in varied, unstructured environments—is the primary bottleneck preventing humanoid robots from becoming useful workers. Existing systems tend to perform well only in narrowly defined conditions, with pre-trained objects, fixed lighting, and carefully controlled environments.

When confronted with novel objects or unfamiliar scenes, many state-of-the-art models fail completely, sometimes achieving 0% success rates. This fragility makes them unsuitable for real-world deployment, where variability is the norm rather than the exception.

General Trajectory is focused precisely on this failure mode: teaching robots not just how to move, but how to generalize manipulation skills beyond what they have seen before.

What Is General Trajectory’s Foundation Model for Humanoid Robots?

At the core of the company’s work is a foundation model designed specifically for humanoid robots performing dexterous manipulation. Rather than training separate policies for each object or task, General Trajectory’s model aims to learn transferable representations of physical interaction.

This foundation model allows robots to pick up unseen objects and perform real-world tasks without requiring object-specific retraining. The emphasis is on generalization—handling unfamiliar shapes, textures, weights, and contexts that were not explicitly present in the training data.

By positioning the system as a foundation model rather than a task-specific controller, General Trajectory aligns itself with the broader trend in AI toward scalable, reusable architectures that improve with more data and deployment.

How Does Reward-Guided Imitation Learning Work in Practice?

General Trajectory’s technical approach centers on reward-guided imitation learning, a hybrid method that combines human demonstrations with learned reward signals.

Instead of collecting massive datasets of exhaustive demonstrations, the company focuses on gathering an efficient and carefully curated set of human examples. These demonstrations show how humans naturally grasp and manipulate objects, capturing subtle behaviors that are difficult to encode manually.

A reward model is then trained to evaluate and improve the robot’s actions relative to these demonstrations. Rather than simply copying human behavior, the system learns to optimize grasps and manipulation strategies by understanding what successful interaction looks like.

This approach allows the base model to refine its performance beyond the initial demonstrations, leading to improved adaptability and robustness when encountering new objects or scenarios.

Why Is Generalization the Key Breakthrough in This Research?

In robotics, generalization is the difference between a lab demo and a deployable system. Many existing models perform well on benchmark objects but collapse when faced with even minor deviations.

General Trajectory’s results highlight the importance of this distinction. On difficult objects where prior state-of-the-art models achieved 0% success, the company reports gains of up to 63%. Crucially, these gains do not come at the expense of performance on standard objects, where the system maintains near-perfect success rates.

This combination—strong generalization without regression—is rare in robotics research. It suggests that the model is learning deeper principles of physical interaction rather than memorizing object-specific heuristics.

Such performance marks a meaningful step toward robots that can operate in homes, factories, laboratories, and field environments without constant retraining.

Why Does Dexterous Manipulation Matter for the Global Economy?

The physical world underpins most economic activity. Manufacturing, agriculture, construction, energy, logistics, and healthcare all rely on human labor interacting with physical objects.

While software automation has transformed digital workflows, physical labor remains largely resistant to automation due to its variability and complexity. Humanoid robots offer a potential solution precisely because they share a similar form factor and interaction space with humans.

However, without reliable manipulation capabilities, humanoid robots remain expensive novelties rather than productive assets. General Trajectory’s work directly targets this economic bottleneck by enabling robots to perform useful tasks across a wide range of environments.

If successful at scale, such systems could address labor shortages, improve safety in hazardous industries, and accelerate scientific research by automating repetitive physical experiments.

How Does This Research Extend Beyond Commercial Robotics?

Although General Trajectory’s immediate focus is humanoid robots, the implications of its research extend far beyond a single product category.

In autonomous defense systems, the ability to manipulate unfamiliar equipment or operate in unstructured environments is critical. In scientific R&D, robots capable of general physical interaction could accelerate experimentation in chemistry, biology, and materials science.

Even outside robotics, insights from reward-guided imitation learning and physical generalization may influence broader AI research, informing how models learn from sparse data and adapt to new domains.

By treating the physical world as a first-class AI environment rather than an edge case, General Trajectory positions itself at the intersection of robotics, machine learning, and applied science.

What Makes General Trajectory Different from Other Robotics Startups?

Many robotics startups focus on vertical integration—building a robot for a specific task or industry. Others concentrate on hardware innovation or mechanical design.

General Trajectory differentiates itself by focusing almost entirely on the learning layer. The company does not present itself as a robot manufacturer, but as a foundational AI company whose models can power a wide range of physical systems.

Its emphasis on generalization, reward modeling, and foundation-level learning places it closer to frontier AI research than traditional robotics engineering. This strategy carries risk, but also the potential for outsized impact if the models become widely adopted across platforms.

The company’s small team size further reflects this focus on deep research rather than rapid scaling.

What Does the Future Look Like for AI in the Physical World?

General Trajectory operates on the belief that the next major wave of AI progress will move beyond screens and servers into factories, labs, and everyday environments.

As digital work becomes increasingly automated, the limiting factor for economic growth may shift to physical labor. AI systems capable of understanding and interacting with the physical world could unlock entirely new forms of productivity.

General Trajectory’s research represents an early step toward that future—a future in which AI systems are not confined to abstract reasoning, but can manipulate, build, and experiment in the real world.

While the company is still at an early stage, its focus on foundational problems positions it as part of a broader movement redefining what artificial intelligence can do.

Why Is General Trajectory a Startup Worth Watching?

With a clear technical focus, a research-driven founder, and a problem that sits at the core of real-world AI deployment, General Trajectory occupies a unique position in the current AI landscape.

Its work addresses a bottleneck that has limited robotics for decades, using modern machine learning techniques to push beyond brittle, task-specific systems.

If the company succeeds in scaling its foundation model and translating research performance into real-world deployments, it could help define the next era of AI—one where intelligence is measured not just by what machines can say, but by what they can physically do.

In that sense, General Trajectory is not just building robots. It is helping AI take its first truly general steps into the physical world.