←
Runhan Huang ✈️ NeurIPS 2026 Atlanta166 posts
Runhan Huang ✈️ NeurIPS 2026 Atlanta avatar

Runhan Huang ✈️ NeurIPS 2026 Atlanta

@RunhanH

PhD student @Harvard | Prev. @Tsinghua_IIIS (Yao Class) | Robot Learning, Generative AI

◷ Joined September 2024▧ 15 Photos and videos
Runhan Huang ✈️ NeurIPS 2026 Atlanta avatar
Runhan Huang ✈️ NeurIPS 2026 Atlanta@RunhanH·Oct 16

Honored to co-lead RAX with Xiangcheng (@xczhangxx)! Excited to share what robots can learn through scaling exploration—no demos, no model finetuning! We’re open-sourcing the full framework and every experiment session. Please fell free to try out! Project + sessions: robot-agent-exploration.com/ Paper: assets.robot-agent-explorati… GitHub: github.com/robot-agent-explo… Huge thanks to @du_yilun for your advice and support, and to all our amazing collaborators—it’s been a joy working with you! @TonyC56348978 @yitongli165665 @wangyenjen @Lyubh22 @haotian_yeee @ChuyeHong @YXWangBot @Huangyu58589918 @Haoyu_Xiong_ @yangzhen04 @xieenze_jr @MandersonTravis @StefanoErmon @pabbeel Antonio Torralba @dawnsongtweets @Weirui_Ye

Robotics has spent years scaling demonstration data and model size. What happens when we scale EXPLORATION for robot agents? We're incredibly excited to share RAX, an agentic robotics system that learns by scaling online exploration. A coding agent connects directly to the robot’s perception and control APIs, learns from feedback, and builds programs that perceive, plan, and act. The model stays frozen. As exploration scales, the agent improves its programs and builds up reusable skills. The results: • ZERO teleoperation or action demonstrations on real robots—from insertion with 0.1 mm clearance and air-fryer manipulation to deformable objects and long-horizon mobile whole-body control. • +37 percentage points on average over leading learned policies across five simulation benchmarks: RoboDojo, BEHAVIOR-1K, HumanoidArena, DexJoCo’s bimanual tasks, and LIBERO-Pro. • One framework spanning single-arm, bimanual, dexterous, humanoid, and mobile robots, with new embodiments connected through APIs. The craziest part: on each real-world embodiment, going from scratch to mastering the tasks took roughly two hours. No teleop. No model finetuning. We’re open-sourcing the full framework AND every experiment session. We can’t wait to see what people build with it! Project + sessions: robot-agent-exploration.com/ Paper: assets.robot-agent-explorati… GitHub: N Project fully led by @RunhanH and @xczhangxx who are absolutely awesome. How it works, experiments, and robot videos below 🧵 (1/N)

2:35
Zhenting Qi avatar
⟳ Runhan Huang ✈️ NeurIPS 2026 Atlanta retweeted
Zhenting Qi@ZhentingQi·Nov 12

Launching the XE² Arena project: a cross-embodiment × cross-environment arena for agentic robots After @claudeai Fable 5.2’s amazing performance on RoboDojo, what's next? In XE² Arena: • Cross-embodiment: one LLM agent, on many robot bodies (YAM, xArm, Wuji Hand 2, humanoid, …) • Cross-environment: every real run gets a sim replica, for both eval and post-training 5 embodiements, 12 real tasks, 400 sim tasks. More bodies & tasks coming soon! xearena.roboticscenter.ai Meet us @RoboticsCtr at CoRL 2026! Joint work with Siyu Lin, @andylizf @yixun_hu @RoboticsCtr @jerryhuang01

1:03
Amani Kiruga avatar
⟳ Runhan Huang ✈️ NeurIPS 2026 Atlanta retweeted
Amani Kiruga@AmaniKiruga·Nov 17

Video models can generate scenes. Geometry models can reconstruct them. What if one model did both? UGG generates video, depth and camera motion together, as one scene. (NeurIPS 2026) Co-first authors: @ZhiyiLi02 and @ruojin8 🌐 ugg-model.github.io

0:55
Runhan Huang ✈️ NeurIPS 2026 Atlanta avatar
Runhan Huang ✈️ NeurIPS 2026 Atlanta@RunhanH·23h

Congratulations, Amani! It's exciting to see RGB and geometry generation brought together in a single unified model!🚀

Video models can generate scenes. Geometry models can reconstruct them. What if one model did both? UGG generates video, depth and camera motion together, as one scene. (NeurIPS 2026) Co-first authors: @ZhiyiLi02 and @ruojin8 🌐 ugg-model.github.io

0:55
Guowei Xu ✈️ NeurIPS 2026 avatar
⟳ Runhan Huang ✈️ NeurIPS 2026 Atlanta retweeted
Guowei Xu ✈️ NeurIPS 2026@Kevin_GuoweiXu·Nov 16

🎉 Excited to share our #NeurIPS2026 paper! 🧠 Do LLMs have a reward system? LLM hidden states are known to encode reward signals like correctness and confidence. But prior work fits black-box probes on the full hidden state, which says little about how this information is organized. We find reward signals live in a small set of neurons: 🔹 Value neurons → predict state value 🔹 Dopamine neurons → encode step-level TD errors

Post attachment
JunLi Ren avatar
⟳ Runhan Huang ✈️ NeurIPS 2026 Atlanta retweeted
JunLi Ren@junli_r84995·Nov 13

What we showed at CoRL 2026: a humanoid with no System 1. An agent-written harness linking System 2 and System 0, plus one human video demonstration, is enough to get the task done. atlas-humanoid.github.io

0:41
Jerry Huang avatar
⟳ Runhan Huang ✈️ NeurIPS 2026 Atlanta retweeted
Jerry Huang@jerryhuang01·Nov 10

Embodied AI: Beyond the Demo. From frontier robot models to real-world deployment. Join us at Robotics Center to connect with researchers and builders advancing physical AI. Details & RSVP:
luma.com/embodied-yac8 #EmbodiedAI #Robotics #PhysicalAI

Runhan Huang ✈️ NeurIPS 2026 Atlanta avatar
Runhan Huang ✈️ NeurIPS 2026 Atlanta@RunhanH·Nov 12

All these paper titles sound like they were named by the same LLM. We may have a benchmark on LLM writing paper titles and add diversity as a metric😂

I scraped all 38,101 ICLR 2027 submissions and counted the names before the colon. LLM has a naming crisis. 🏆 The leaderboard: TRACE: 47 papers PRISM: 42 PACE: 22 SCOPE: 21 SAGE: 17 MOSAIC: 16 COMPASS: 14 750 names are used by 2+ papers. That's 2,209 papers sharing a name. Best parts: • Two papers called TRACE both fix VLM hallucination. Same name, same problem, different authors. • Four papers called PRISM do time series. Good luck, reviewers. • "PACE: Discovering Options via Self-Distillation." None of P, A, C, E appear in the title. They picked the name first. • 47 TRACEs cover quantum computing, materials discovery, seizure localization, carbon emissions, traffic control, and circuit design. TRACE is everywhere. • The formula: 5 letters, looks like a real word, sounds cool. CAST. PACT. CRAFT. GRAFT. FORGE. Tip for ICLR 2028: search OpenReview before you name your method 😂 Follow for more data-backed ML nonsense.

Runhan Huang ✈️ NeurIPS 2026 Atlanta avatar
Runhan Huang ✈️ NeurIPS 2026 Atlanta@RunhanH·Nov 12

4/4 Better takeover decisions. Better learning. 🚀 With matched data and training budgets, mean autonomous success across 3 runs: Block stacking EqC 96.7% · Diff-DAgger 80.0% · human-triggered intervention 58.3% Garment folding EqC 98.3% · Diff-DAgger 81.7% · human-triggered intervention 55.0% Energy-guided intervention beats both learned failure detection and human judgment at selecting recovery data that improves the policy.

Post attachment
Runhan Huang ✈️ NeurIPS 2026 Atlanta avatar
Runhan Huang ✈️ NeurIPS 2026 Atlanta@RunhanH·Nov 12

3/4 Turn “I need help” into “I can do it.” 🔄 When EqC detects persistent inconsistency, it pauses and calls Astra. The agent adapts its prepared robot skills to recover; those recovery trajectories train the next VLA. Detect → agent recovery → policy update → greater autonomy. Energy-Gated DAgger × Astra turns agent expertise into lasting robot capabilities: a practical recipe for lifelong learning and RSI.

Post attachment
Runhan Huang ✈️ NeurIPS 2026 Atlanta avatar
Runhan Huang ✈️ NeurIPS 2026 Atlanta@RunhanH·Nov 12

2/4 How does a policy know when to ask for help? Energy consistency. We compare energy progress along the action-refinement trajectory with the relation learned from demonstrations. Persistent deviations trigger a request for assistance. Introspection comes directly from the controller’s own optimization trace—no extra uncertainty model or expert action needed at inference.

Post attachment
Runhan Huang ✈️ NeurIPS 2026 Atlanta avatar
Runhan Huang ✈️ NeurIPS 2026 Atlanta@RunhanH·Nov 12

1/4 How does EqC work? 🧠 Inspired by Equilibrium Matching @du_yilun, we learn an explicit conditional energy landscape over action chunks. Its gradient becomes the control field: noise → energy-gradient refinement → actions. The formulation works with both diffusion-transformer policies and flow-matching VLAs. The same landscape that generates actions also reveals when the policy is reaching its limits.

Post attachment
Runhan Huang ✈️ NeurIPS 2026 Atlanta avatar
Runhan Huang ✈️ NeurIPS 2026 Atlanta@RunhanH·Nov 12

What should control policies and VLAs look like in the era of agentic robotics? Policies (e.g., VLAs and WAMs) should know their capability limits, call an agent for help, and turn every intervention into greater autonomy. Introducing Equilibrium Energy Control (EqC): control with introspection, built for self-improving robots. 🤖 🧠 One energy landscape. Both action and introspection. Using an energy-based action generator, EqC generates actions while monitoring its own competence through the geometry of action refinement. The policy itself knows when to ask for help. 🔄 Energy-Gated DAgger × Astra: an agent-powered self-improvement loop. Detect a competence gap → request agent recovery → learn from it → expand autonomy. A concrete recipe for RSI in VLAs. More effective than human judged intervention. • Real-world garment folding: 0% → 100% autonomous success after just 6 Astra-generated recoveries. • Real-world block stacking: 0% → 96.7%. • Real-world debris cleaning: 100% success. 🚀 Stronger control across four benchmarks. EqC boosts π₀ on RoboTwin 2.0 Hard (18% → 33%) and LIBERO-Plus (53.6% → 65.8%) in a plug-and-play manner! Accepted as a Spotlight at the Continually Self-Improving Robots (CSIR) Workshop @ CoRL 2026! Come see the poster—and join for my spotlight talk: 🗓 Nov 12, 9:25–9:30 AM CT 📍 Room 202, JW Marriott Austin Code is open source! 📄 Paper: arxiv.org/abs/2610.18790 🌐 Project & demos: equilibrium-energy-control.g… 💻 GitHub: github.com/hrh6666/EqC #RobotLearning #AI #LLM #Agents #RSI #corl2026 #corl #VLA #Control #Robotics #EBM

3:01
Guowei Xu ✈️ NeurIPS 2026 avatar
⟳ Runhan Huang ✈️ NeurIPS 2026 Atlanta retweeted
Guowei Xu ✈️ NeurIPS 2026@Kevin_GuoweiXu·Nov 10

🚀 Can AI scientists come up with truly creative ideas? A creative idea needs more than good performance. It must be original, useful and non-obvious. Yet AI scientists are usually given as much task information as possible, which anchors them on familiar approaches. We introduce BVSR: Blind Variation, Selective Retention, a discovery framework inspired by a classic theory of human creativity. ✅ Plugs into existing AI-scientist workflows.

0:54