Beatposts @Beatposts - Twitter Profile

Pinned Tweet

11 months ago

@newstart_2024 100% agree. Its all fake shit on computers. People still cook same way, travel same way but worse , build same but worse , sleep & dress same but worse. There has been zero development only pimp like business goons extracting money. I feel like AI is same packaging of same crap.

1

26

4

1

1K

Beatposts retweeted

FHILY👑

@Oluwaphilemon1

about 2 hours ago

An open 27B Qwen model has reportedly learned to control robot arms through its own attempts in simulation, without being trained on human demonstrations. The project, PEAR from MIT HAN Lab, builds on Qwen3.8-27B and takes a different approach to teaching AI systems how to interact with the physical world. Instead of showing the model thousands of examples of how a task should be performed, the researchers trained it through trial and error in a simulated environment. It attempts tasks, observes what happens and learns from its own experience. The result is a model that can move beyond generating instructions and actually translate them into physical actions. From simulation to a real robot The most interesting result is its performance on a physical robot it had never interacted with before. Before training, it completed 0 out of 30 tests. After training, it succeeded in 18 out of 30, giving it a 60% success rate in the reported evaluation. That is a meaningful improvement because controlling an unfamiliar robot requires more than understanding a verbal instruction. The model has to interpret its surroundings, determine what action to take and coordinate movements to achieve the intended result. Its reported capabilities also extend beyond basic object manipulation. The model can: Reproduce a painting through robotic interaction. Design its own chemistry experiment. Follow instructions for unfamiliar tasks. Transfer skills learned in simulation to a physical robot. These examples suggest that the training is developing more general interaction skills rather than teaching the model to repeat a fixed sequence of movements. Why learning without demonstrations matters Most imitation-learning approaches depend on demonstrations. A human or another system performs a task, and the model learns to reproduce those actions. This works, but collecting demonstrations can be expensive and time-consuming. Every new task may require additional examples, especially when the environment or robot changes. PEAR explores a different route The model learns through its own attempts in simulation, without relying on human demonstrations as the training source This could make it easier to develop skills across a wider range of tasks, although simulation-to-reality transfer remains a difficult problem. Physical robots encounter differences in friction, object weight, positioning, sensor readings and movement that simulations may not perfectly reproduce The reported real-robot results are therefore particularly important How does it compare with other models? According to the researchers, PEAR generalises and follows instructions better than π0.5 On LIBERO, a benchmark for evaluating robot manipulation and task execution, its performance comes close to GPT-6 Astra Max These are reported comparisons, so the evaluation conditions and exact scores are important when interpreting the results Another notable detail is that the robotics training reportedly preserves the model’s general capabilities Math and coding performance remain intact, rather than being sacrificed to specialise the model for robot control That matters because a useful robotics model needs more than movement skills. It also needs reasoning, planning and the ability to understand complex instructions It can run locally The model is built on Qwen3.8-27B and reportedly generates at 220+ tokens per second on an NVIDIA RTX 5090 That makes the project particularly interesting for developers exploring locally deployed robotics systems Rather than depending entirely on a large proprietary model in the cloud, researchers can experiment with an open-weight model and adapt its capabilities to their own hardware and workflows The researchers have also released the model weights, code and full training infrastructure That gives other developers an opportunity to inspect the approach, reproduce the experiments and explore further applications

1

3

2

5

1K

Beatposts

@Beatposts

Last Seen Users on Sotwe

Trends for you

Most Popular Users