STASHERiLiAS @STASHERILIM - Twitter Profile

Robora Development Update At Robora, we take huge pride in keeping our community informed as we build. Every step forward is not just a technical milestone but a shared achievement, and that’s why we’re excited to share these updates with you: 1. Model Weights Management Added support for pulling open-source model weights directly from Hugging Face via the huggingface-hub Python package. Built an early CLI tool with rich terminal feedback to simplify model management. 2. Fine-Tuning SDK Roadmap Defined the path for our fine-tuning SDK: starting with Imitation Learning (IL) to adapt VLA models to new robot morphologies. Next, RLHF fine-tuning inside PyBullet, enabling models to handle new tasks, even adapting to edge cases like a robot continuing after losing functionality (e.g. a “broken leg”). Chose PyBullet as the simulation backend for its CUDA acceleration and reliability (with all configs stored in sims-env). 3. 3D Mapping SDK (In Progress) Began development of a companion SDK for creating 3D point cloud maps of real environments. Currently experimenting with photogrammetry to convert raw camera input into 3D meshes, aiming to reduce reliance on LiDAR. 4. Documentation Added roadmap direction and high-level guidance for contributors and interested developers in the Docs directory. 5. Initial Model Integration Prioritized early support for SmolVLA (efficient) and Pi0 (generalist policy) as part of the VLA SDK integration. 6. Development Priorities Current focus: fine-tuning in PyBullet and generating rich 3D simulation datasets. Next steps: inference logic and cloud/online integration, either in parallel or after the simulation-focused phase.