Vulcan Dynamics — ten fine-tuned ACT specialists
Selected controller: ACT_ONLY, H50 for all ten official tasks. Each models/R5/TASK directory contains model.safetensors, config.json, train_config.json, finetuning_provenance.json and an individual README. MODEL_MANIFEST.json maps the tasks and immutable official parent revisions. CHECKSUMS.json records file sizes and SHA-256 values. Normalization means and standard deviations are embedded in model.safetensors; no external normalization file is required by the pinned official ACT loader.
These are genuine supervised fine-tuned derivatives of task-specific RoboSynChallenge ACT parents, using their official simulation demonstrations. They were not trained from scratch. Nine specialists received 100 participant optimizer updates; Water Pouring received 200. Recorded R5 settings: AdamW, learning rate 1e-5, batch size 4, weight decay 1e-4, gradient clipping 10, seed 1037, 64 training and 16 validation episodes. Consult each immutable finetuning_provenance.json and the pinned task registry for exact lineage. train_config.json describes historical upstream training and is not our training recipe; historical paths were redacted for publication.
Architecture: LeRobot ACT, three camera inputs plus 14-dimensional robot state, 14-dimensional actions, 50-action queue, reset at each episode. The official pretrained backbone, preprocessing, configurations and normalization remain in use. Our contribution is task-specific supervised fine-tuning, specialist routing, provenance, reproducible packaging and simulator-based engineering evaluation.
R5–R10 contain genuine CUDA loading and simulator evidence. All ten ACT specialists executed simulator rollouts in earlier stages. R10 verified all ten ACT routes through CUDA loading and relocated-path loading; its hybrid integration episode for Item Assembly used experimental DP and must not be presented as ACT validation. Experimental DP is excluded here. No GPU or simulator experiment was repeated in R11. Small cohorts, selection effects, poor performance on several tasks and intermittent native simulator startup faults limit interpretation. No statistically established improvement over released ACT parents, physical-robot validation, universal success or competition acceptance is claimed.
Attribution: RoboSynChallenge/EDEM-AI; LeRobot/Hugging Face; ACT by Zhao et al., Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware; official task demonstrations and parent checkpoints individually linked in the manifest. See LICENSING.md: upstream checkpoint redistribution permission and competition eligibility remain unresolved. No model license is inferred from Apache-2.0 source licensing.