Tsinghua open-sources RPent, an agent layer for real robots
Two open-source releases today point the same direction: the interesting work is no longer in the model itself, but in the scaffolding around it. A Tsinghua-led team published the layer that sits between a frontier model and a robot's motors — and a leaderboard to show it lifts a frozen policy. A solo developer, meanwhile, trained a small continual-learning model on a laptop GPU and documented what it forgets.
A Tsinghua-led group open-sourced RPent, a framework that turns a general model into the planner and leaves the actual manipulation to specialist policies. The project is a joint effort from Tsinghua's team led by Chao Yu, Infinigence (无问芯穹) and Zhengxing Innovation, built on the RLinf embodied-RL stack, and it lands with an interactive leaderboard across four robot benchmarks. RPent's bet is that a general model should not output joint commands. It should decompose a task, call tools, check results and retry — while a frozen vision-language-action model handles the sub-centimetre grasping it was trained for.
That split is the whole design. The planner is swappable: RPent ships adapters for Claude Code and Codex, and accepts a custom planner. Action primitives wrap VLAs such as Pi0.5, RLDX-1 and LingBot-VLA, plus programmed skills, behind one interface, so the same task logic runs on a simulator or a physical arm. Today's release covers LIBERO-PRO, RoboCasa, RoboTwin and RoboDojo in simulation, and Franka, dual Franka, YAM and SO-101 arms in the real world.
The part worth watching is memory. RPent stores Recipes — validated task flows with check conditions, deliberately without hard-coded coordinates — and compresses successful runs into Task Cards. In Flash Mode the agent replays a card, re-locating objects visually, and only wakes the planner when a check fails or the scene drifts. The team reports the non-reasoning path cuts average execution time by about 40%, which is the difference between a demo and something that finishes before the shift ends.
The numbers are the argument. With Codex driving and GPT-6 Astra underneath, RPent reports 92.63% overall (741 of 800) across all eight LIBERO-PRO suites. Swapping the planner changes the ceiling: Opus-4.7 reaches 96.0% on LIBERO, GPT-5.5 takes 62.4% on RoboTwin, and on RoboCasa365 Target50, RPent with GPT-6 Astra scores 59.20% against 57.1% for the same harness on GPT-5.5. Read that second pair carefully — a better base model bought about two points on a long-horizon kitchen task. The harness is doing more of the work than the upgrade.
Separately, a Show HN project called mini-AGI trained a continual-learning language model from scratch on an 8GB laptop GPU, one batch at a time. Its author, volotat, set out to answer a narrow complaint: every model you can own is somebody else's frozen weights, and fine-tuning on your own data makes it forget. The architecture stores each mixture-of-experts expert as a file on disk, pages a working set in and out based on what the incoming text needs, and grows new experts only when every guard agrees — a new expert is protected for a full survival window so it cannot be pruned before it has been tried. Positions are rotary and carry no learned parameters, so the context window extends by continuing to train rather than by re-initialising.
The honest part is the forgetting curve. Reading 500,000 characters of chess at the experts' own learning rate pushes seven unrelated subjects from 1.12 to 3.73 nats, and the project's own write-up corrects an earlier account of its own results. The run reads a 7.87-billion-character corpus in a single pass over days to weeks, with code, chat and reasoning improving fastest while Wikipedia stays the expensive domain. It is early — 35 stars, days old, and the author discloses that Claude Opus 5 wrote most of the code — but the retention gap between 50.68% and 99.84% is reported as one setting in a config file, which is exactly the kind of knob the field needs more of.
What to watch: whether RPent's leaderboard holds up when independent labs run it on their own hardware, and whether the harness-over-model result shows up in arms people actually buy.
Would you let a planner model — not the arm's own policy — decide your next grasp? Tell us in the comments.
Sources: RPent (GitHub) · mini-AGI (GitHub)
