.png&w=256&q=75)
🤝 Top Collaborators
.png&w=640&q=75)
.png&w=640&q=75)
🤓 Latest Submissions
Berkelium Neural Rendering Studio
Berkelium Studio is an open, AI-native desktop studio for 3D creation that connects natural language, real engineering geometry and neural rendering. Type "a pipe flange with a 53 mm bore, 290 mm outer diameter and 3 bolt holes" and Studio returns an exact, measurable part, not a guess. The core idea is a strict separation of roles: the AI proposes and deterministic code decides. A fine-tuned language model turns the request into a parametric design: either a Berkelium DesignProposal (named dimensions, requirements, assumptions and open questions) or a CadQuery script. The Berkelium engineering core then builds exact geometry with OCCT and Manifold, measures it and validates it against every stated requirement. Generated code runs in a locked-down sandbox. Finally, Studio fits Gaussians to the result and splat-renders it in the viewport. Rendering uses AMD's gsplat with our own deterministic HIP backward pass, verified on an MI300X with bit-identical gradients and a 1.3–2.2× faster backward pass. We trained two models. BerkeliumGPT-8B, a Qwen3-8B LoRA fine-tuned on Kaggle T4s using 18.7k core-verified examples, builds 99.8% of designs for its trained part families. 99.7% meet every stated size, with an IoU of 0.94 against the reference designs. Base Qwen3-8B scores 0% on the same prompts. BerkeliumGPT-32B, trained on an AMD MI300X with Unsloth on ROCm, writes CadQuery from 121k sandbox-verified examples covering washers to rocket airframes. Both are public on Hugging Face, with GGUF builds runnable through Ollama. Every promotion rule was pre-registered before training, and every result, including models that were not promoted, is published with hashes and receipts. The open-world text-to-3D generators (Shap-E, Z-Image → TripoSR) are benchmarked on 52 prompts so that evidence, not hype, picks the defaults. Berkelium is engineering-grade generative design you can verify.
18 Oct 2026
Bimanual Pick-Handoff-Place with VLM Planning
Two simulated SO-101 arms face each other across a table. Arm A picks up a cup, hands it to Arm B across a shared reach zone, and Arm B places it on a marked target — all driven by a natural-language instruction that a hosted vision-language model (Gemini 3.6 Flash) interprets alongside a rendered camera frame to select which scripted primitive to run next. A local CLIP model, converted to OpenVINO IR and benchmarked on Intel CPU, cross-checks that the target object is actually visible in the scene, giving the submission a genuine local-inference component alongside the hosted VLM. All arm motion uses classical inverse kinematics and waypoint control rather than a learned policy — a deliberate scope decision made under a tight timeline, documented openly in the repo's README. The system was evaluated across 10 randomized seeds (varying cup position, color, and lighting) with a 100% task success rate. Several episodes exercised the planner's fallback path after hitting the VLM provider's free-tier rate limit, demonstrating graceful degradation rather than failure. The OpenVINO benchmark shows FP16 CLIP inference running about 11% faster than FP32 on the available hardware, with half the model size. As an online-track team, we benchmarked on our own available Intel Core i7 hardware rather than Core Ultra Series 2/3, which wasn't provided to remote participants — this and every other simplification (kinematic grasping, position-only IK, narrow randomization) is documented transparently in the repository rather than glossed over.
16 Sep 2026
