Important
Prizes: Up to $10,000 USD for top-tier AI Agent builders in Slack! Let's redefine collaboration! 💬🚀
Forenly AI Systems' entry for the Slack Agent Builder Challenge (Salesforce / Slack) — bringing Reinforcement Learning from Human Feedback (RLHF) and active humanoid skill training into Slack.
Forenly AI Skill Agent is the human-in-the-loop (HITL) control center for humanoid robot skill acquisition. Instead of requiring complex developer terminals or standalone ML dashboards, AI researchers and robot operators can monitor active training, score humanoid training episodes to provide human preference rewards (RLHF), and guide active reinforcement learning agents through complex physical scenarios—all directly from Slack.
Building this in the open — join the team chat on Discord: https://discord.gg/DCmAYYE32h
- Slack AI capabilities — natural-language queries to fetch real-time training diagnostics, reward curves, and loss trends.
- MCP server integration — Forenly's custom Model Context Protocol (MCP) server connects Slack directly to the active MuJoCo simulation, enabling real-time training state retrieval and remote intervention.
- Real-Time Search (RTS) API — search and index thousands of simulated training episode logs and anomaly reports inline in Slack.
Training humanoid robots (like the Unitree G1) in physics simulators (MuJoCo) via Reinforcement Learning (RL) suffers from the sparse reward problem—it is hard to mathematically define "natural walking" or "graceful grasping."
Forenly AI Skill Agent solves this by putting human operators directly in the training loop:
- Interactive RLHF Reward Scoring: The robot performs trial runs in MuJoCo, renders short 5-second video clips of its performance, and pushes them to Slack as a Block Kit Skill Audit Card. Operators rate the run (1 to 5 stars) directly in Slack, instantly writing human preference scores back into the model's reward function.
- Active Training Intervention: When an RL training run encounters a critical bottleneck (e.g., reward flatlines or eklem motor torque overshoots safety limits), the simulation pauses. The agent posts an alert card to Slack, allowing operators to reset to the last stable checkpoint, inject manual rewards, or hot-swap active LoRA adapters with one click.
- Conversational Training Queries: Ask Slack AI about live training progress: "Summarize G1_0's walk-slippery-floor learning success rate over the last hour" or "What's the active loss value of the bimanual_manipulation run?"
See ARCHITECTURE.md for the detailed closed-loop humanoid training architecture.
| Date | Event |
|---|---|
| 14 Jul 2026, 01:00 GMT+1 | Submission deadline |
- Track selected (New Slack Agent)
- Text description (features + functionality)
- Demo video ~3 min, shows the working project
- Architecture diagram
- URL to Slack developer sandbox — grant access to
slackhack@salesforce.comandtesting@devpost.com
Technological Implementation (uses ≥1 of Slack AI / MCP / RTS) · Design (UX; FE/BE balance) · Potential Impact · Quality of the Idea.
Per track (New Agent / for Good / for Orgs): 1st $8,000, 2nd $4,000 (+ Dreamforce 2026 vouchers, certification, swag for 1st). Plus Best UX $2,000, Most Innovative $2,000, Best Technological Implementation $2,000.
🚧 Pre-build. Track locked (New Slack Agent). Concept updated and locked (Humanoid RLHF & Active Skill Training).
Forenly AI Systems · github.com/Forenly