You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
felt-agent (github.com/LukeHamond1001/felt-agent) is my open-source architecture for embodied agents that generate their own reward from a grounded appraisal, rather than being handed one. Two of its mechanisms are directly safety-relevant and already validated in simulation: anti-wireheading by construction (the drive is the prediction error over grounded reward — ablation-tested, removing it triples wireheading rate) and structural real/imagined separation (imagined states carry a hard provenance bit below any learned component, so the action path physically cannot act on imagination). This project moves both to hardware for the first time, inside the regime safety research most needs data on: a continually-learning agent. One Dreamer 4-style transformer world model (vision + audio + proprioception fused in a single latent, grounded reward head seeded from my measured zero-shot affect-grounding result, AUC 0.95), pretrained offline on demonstrations collected by physically moving a leader arm (LeRobot SO-101, $500), then deployed for 24/7 continual RL on a local rig — measuring where grounding, anti-wireheading, and retention hold or break.
Goal: a public, measured answer to whether a modern world-model agent with a self-generated grounded reward can keep learning on real hardware — and exactly how its safety-relevant mechanisms behave while it does. 16 weeks: (1) weeks 1–3, build the SO-101 teleop rig and collect large-scale multimodal demonstrations; (2) weeks 4–8, pretrain the unified world model offline and train the policy inside its imagination, Dreamer 4-style; (3) weeks 9–13, deploy for 24/7 continual RL on a local rig; (4) weeks 14–16, measure and publish the failure modes everyone names and nobody quantifies on hardware — plasticity loss, catastrophic forgetting, reward-grounding drift as the representation shifts — plus verification of the provenance-bit interlock and RPE-habituation (anti-wireheading) on the physical system. Everything ships open: code, data, and the failure report.
Total goal $15,000: SO-101 leader/follower arms + cameras + microphones ($1,200); used RTX 4090 workstation to run the 24/7 continual-learning loop locally — cloud costs for an always-on loop are ruinous ($2,500); cloud GPU burst for the offline multimodal pretrain ($3,500); parts, task objects, storage ($800); ~16-week part-time stipend so I can execute alongside undergraduate coursework ($3,500); contingency — the official Dreamer 4 code is unreleased, I'm building on an unofficial implementation and budgeting debugging time ($3,500). Minimum funding $4,000 covers hardware + the local rig with no stipend: slower, but the loop still runs and everything still ships.
Solo: I'm Luke Hamond, 20, studying electrical engineering online through North Dakota State University — no lab, no advisor, self-taught in ML/RL. Track record: I designed, built, and published felt-agent over the past year — a ~70,000-word open engineering spec plus running experiments, including a measured zero-shot affect-grounding result (AUC 0.95 from frozen features) and simulation validation of the anti-wireheading mechanism. The repo labels every component "proven" or "bet," and when my cheapest test of the core thesis failed, I pre-registered a test of my own excuse and published that it failed too — I report negative results, funded or not.
Most likely failure: the continual-learning loop degrades — plasticity loss or catastrophic forgetting collapses performance, or the grounded reward drifts as the representation shifts. That outcome is itself the deliverable: a measured, public account of where and how it breaks is more useful to the field than a fragile success. Second risk: the unofficial Dreamer 4 codebase costs more debugging time than budgeted (mitigated by contingency, and by falling back to the proven DreamerV3 recipe with the same measurement plan). Third: hardware fragility on a $500 arm (mitigated by spares). Worst case across all three: fewer tasks and a narrower failure report, still published with code and data.
$0. This would be the project's first outside funding. An Emergent Ventures application for the same project was submitted this week, and I'm applying for free compute via the TPU Research Cloud; I'll update this post if either lands.