Building the Foundation for Real-World 3D Capture We’re making strong progress on Robora’s next step: enabling people to easily capture real-world 3D data using affordable hardware. 1. Smarter Camera Setup We’re finalizing a simple multi-camera system that works together with a depth sensor. This setup allows anyone to record high-quality, multi-angle scenes without expensive gear, making 3D data collection more accessible. 2. Hybrid Processing Pipeline Captured footage is lightly processed on the device (to blur faces, remove license plates, and protect privacy) before being uploaded. The heavy lifting (turning that footage into detailed 3D environments) happens on powerful cloud servers using advanced reconstruction methods. This design keeps the hardware simple while still producing professional-grade visual results. 3. Data Protection & Transparency We’re also adding a compliance layer to ensure all collected data meets privacy and legal standards. Contributors will know exactly how their data is used, how long it’s stored, and what rights or royalties they hold over it. Why This Matters This setup lays the groundwork for Robora’s ecosystem, where creators, developers, and robotics teams can capture and share real-world 3D data safely and efficiently. It’s a major step toward large-scale, decentralized 3D scene generation. What’s Next We’ll begin real-world testing of the capture kit and cloud reconstruction pipeline, fine-tuning the process before opening it up to early contributors.

We’ve officially decided to keep the buy tax to 0% indefinitely to support our community and encourage continued growth. The sell tax remains at 4% as normal, helping sustain the project and its future developments.

Buy and sell tax is now 0%. This decision follows detailed financial modeling confirming our ability to sustain operations. New revenue streams are being developed to ensure long term stability. Removing the tax aligns incentives across holders, traders and partners as we move toward a more organic, utility-driven growth model.

1. Objective This experiment was designed as a proof of concept to show that our VLA model (Pi0) can be successfully fine-tuned for robotic tasks and that our training pipeline, including Weights & Biases (W&B) tracking, works as intended. This integration enables comprehensive tracking of training metrics, hyperparameters, and system utilization, helping ensure better reproducibility and optimization in future fine-tuning runs. 2. Why This Matters This run proved that: The Pi0 VLA model learns properly on real robotic data. W&B analytics integration works, giving us full visibility into loss curves, gradients, GPU usage, and training behavior. We now have a baseline VLA fine-tuning setup that can be repeated and improved. For more technical information, please read on👇 A. Experiment Setup A fine-tuning session was conducted on the Pi0 model, known for its strong embodiment and visuomotor grounding capabilities. The training used the SOARM101 dataset and was executed on an NVIDIA H200 GPU. Key configuration details: - Model: Pi0 - Paligemma backbone (3B) + Gemma based action expert (300M) - Dataset: SOARM101 (robotic manipulation tasks) for POC - Training Steps: 20,000 - Peak VRAM Utilization: Around 35 GiB - Optimizer: AdamW - Learning Rate Scheduler: CosineAnnealing - Framework: Robora VLA fine-tuning pipeline B. Parameter-Efficient Fine-Tuning (PEFT) Integration In addition to the baseline run, follow-up experiments are planned using PEFT techniques such as LoRA and QLoRA. These will evaluate trade-offs between VRAM efficiency and fine-tuning performance. The goal is to establish optimal configurations for efficient adaptation of embodied vision-language-action (VLA) models under limited compute conditions. C. Observations and Analytics The W&B analytics provided detailed visibility into multiple training aspects, including: - Loss convergence across 20K steps - Gradient stability and magnitude distribution - Learning rate dynamics under cosine annealing LR Scheduler wrapper. - GPU utilization and memory efficiency - Optimizer–scheduler interaction effects on overall training performance These analytics will guide the selection of optimal optimizer–scheduler pairs and hyperparameter configurations for upcoming fine-tuning cycles. D. Conclusion and Next Steps This session successfully validated W&B analytics integration and established a strong baseline for Pi0 model fine-tuning on the SO-ARM101 dataset as a proof-of-concept. As soon as our Sim-data pipeline will be ready, we will be able to run even more robust and good fine-tuning session for different robot morphology and tasks. The next phase will include comparative evaluations using LoRA and QLoRA to analyze efficiency and performance trade-offs. Future experiments will also track inference latency, task completion accuracy, and embodied control performance metrics.

We’re closing another strong month at Robora, one filled with real progress, new collaborations and innovation. To everyone who’s been part of this journey, thank you. Your support drives the mission of building a verifiable, open future for robotics. Together, we’re showing how far a community can go when driven by a shared mission. Here’s a look at what we’ve been working on this October, in our latest Medium article👇 robora.medium.com/monthly-re…

VLA Fine-tuning Module Update A. Fine-Tuning Pipeline Enhancements - Added default configurations and state management for multiple optimizers, including SGD, SGD with Momentum, Adam, and AdamW, enabling end-users to experiment with different optimization strategies. github.com/RoboraDev/VLA_Mod… - Integrated several learning rate scheduler wrappers -> StepLR, CosineAnnealing, LinearDecay, and ExponentialDecay, allowing seamless selection and comparison through Wandb for fine-tuning analytics. github.com/RoboraDev/VLA_Mod… - Default implementations for Pi0, Pi0.5, and SmolVLA support a maximum action and observation dimension of 32. For high-complexity robotic agents such as Humanoid, the action encoder, state encoder, and action decoder feature dimensions will need to be scaled accordingly, these can be done easily as these are 3 lines of pytorch code but have to be implemented from scratch. For each model, this will be specified inside config.py and the corresponding WithExpert.py file will inherit it. github.com/RoboraDev/VLA_Mod… B. PI0 Policy Implementation - Implemented the Pi0 Policy architecture in PyTorch, referencing the Physical Intelligence OpenPI repository for core design principles & configurations. - Implemented PI0Config, will also be added yaml, json support for custom configs in fine-tuning. github.com/RoboraDev/VLA_Mod… C. VLA Util helper functions Implementation - get_device_info, torch_device, device_name for accelerators - JSON Serialize & deserialize functions to store the optimizers state on disk. - Parameters utility helpers : get device, dtype, output shape etc github.com/RoboraDev/VLA_Mod… Plan for Tomorrow: - Support for SmolVLA and Pi0.5 will be implemented next, as these architectures share a common Vision-Language Model (VLM) architecture with minor variations in module structure and feature dimensions. - Lerobot dataset framework integration for dataset management: will integrate online dataset as a Proof-of-concept.

This week, we’ll begin pilot deployments of the capture pipeline within our simulation environment, an intermediary step before scaling to real-world data collection. These tests will validate the full end-to-end flow: from visual capture and 3D reconstruction, to data upload, fine-tuning and behavioral analysis of the VLA model. The purpose is clear: to evaluate how well our models adapt when exposed to semi-synthetic, simulation-anchored data that mimics real-world complexity. By introducing variability in lighting, geometry and object dynamics, we can measure the model’s domain adaptation efficiency and its capacity to generalize across unseen conditions, which is a key metric for narrowing the sim-to-real gap. We’ll also launch the first fine-tuning experiments through the VLA SDK, processing captured scenes and quantifying shifts in model behavior compared to purely simulated inputs. This will help us refine both our data pipeline and our fine-tuning methodology before scaling up to large-scale physical capture campaigns. In parallel, work begins on the Robora Data Policy & Contributor Incentive Layer, designed to credit and reward those who supply valuable visual data. This framework will connect to Robora’s on-chain architecture, establishing verifiable proof-of-contribution and preparing for tokenized incentive mechanisms. The ultimate goal: to transform data collection from a passive process into an active, community-driven ecosystem where every user helps teach embodied intelligence to see, understand and evolve.

Highlights this week: • Lemorele P300 (R) Integration The Lemorele P300 (R) officially joins Robora’s hardware suite, turning every user into a real-world data contributor. It enables real-time, high-definition video capture from any camera (robot, drone, or handheld) and wirelessly streams it to devices running the Robora VLA interface or data capture app. These feeds are then uploaded directly into Robora’s cloud or local VLA nodes, powering real-world model training and fine-tuning. • Smart Contract Audits Completed We finalized our DApp and successfully conducted audits on the smart contracts that will be deployed, through @SolidProof_io . This solidifies the foundation for our upcoming releases. • VLA Fine-Tuning SDK Pipeline Implemented the full code pipeline of the VLA Fine-Tuning SDK, connecting live capture, processing, and model fine-tuning into a single, automated flow. • 3D Reconstruction & Scene Understanding Advanced work on building complete 3D scenes from raw visual data, reconstructing geometry, texture, and spatial layout for real-world mapping and model sharing. On top of that, scene-understanding algorithms now analyze these reconstructions for object segmentation, semantic labeling, pose estimation, and environment profiling, creating structured data that robots can actually learn from. This dual process (reconstruction + understanding) is grounded in the latest robotics research showing that blending real-scene 3D data with simulation sharply reduces the sim-to-real performance gap, enabling more robust perception and adaptability in embodied AI systems.

Real-world data is the missing link between simulation and physical intelligence. With the Lemorele P300 (R), Robora is closing that gap, turning every user into a contributor to embodied AI evolution. It enables real-time, high-definition video capture from any camera, whether it’s mounted on a robot prototype, a drone, or a handheld camera. This visual feed is transmitted wirelessly and losslessly to a connected device (tablet, phone, or PC) that runs the Robora VLA interface or data capture app. The device then streams or uploads these feeds directly into Robora’s cloud or local VLA processing node. By collecting data from diverse environments and use cases, Robora gains the foundation to train and fine-tune its models on real-world sensory information, moving beyond purely simulated data. This approach is key to reducing the sim-to-real gap, the performance difference between robots trained in simulation and those operating in complex physical environments.

This week, Robora made solid progress on multiple fronts as we continued to refine our robotics stack, improve performance and move closer to real-world deployment. Our focus has been on improving model adaptability, advancing physical intelligence and strengthening the hardware foundation that supports our platform. We introduced Ege, an industrial designer from Istanbul who plays a key role in shaping Robora’s hardware systems. His background in aviation and product design helps bridge creativity and engineering, designing structural systems that bring our robots to life both in simulation and in real-world environments. Significant strides were also made in hardware integration. We successfully addressed three out of five major challenges for vision-to-prompt integration, including signal communication, wireless connectivity and onboard SDK support. These improvements bring us closer to a fully functional hardware layer that connects directly with our software stack. On the development side, two key components of our Physical AI architecture saw major updates. The first is the VLA Fine-Tuning and Adaptation Pipeline, which enables our models to support different robotic hardware configurations while preserving their reasoning capabilities. Using efficient fine-tuning techniques like QLoRA, we can retrain only the action layer, allowing a single VLA model to adapt to new modules such as grippers, arms or mobility units. In parallel, we continued building a PyBullet-powered simulation environment for reinforcement learning. This platform trains low-level control policies for locomotion, balance and stability in complex and dynamic conditions, using advanced algorithms like PPO and SAC. Our approach relies on a clear separation of control: the VLA model serves as the high-level planner, understanding visual input and task context, while RL-based controllers handle precise, real-time execution. This layered design allows our robots to combine intelligence with physical robustness and adapt to a wide range of environments and mechanical configurations.