Robot Era (星动纪元)
Tsinghua-spun embodied-AI company whose open-source VPP2 world-action model topped the RoboDojo benchmark, beating GPT-6-Astra and π0.5.
1. Core Product / Service
Robot Era builds humanoid and mobile robots plus the "brain" that controls them: STAR1 and Star L7 (full-size bipeds), Q5 (wheeled service robot), XHAND1 (five-fingered dexterous hand), and ERA-42, its native end-to-end embodied-AI model [1]. Hardware is reported >95% self-developed [local].
Its headline software is the VPP (Video Prediction Policy) line: VPP2, a World Action Model (WAM) announced ~Oct 2026, learns manipulation by predicting future video frames and then distills the policy into a small action model. VPP2 is 14B parameters, fully open source, trained in three stages on Alibaba's Wan2.1-I2V-14B video model, and tops the RoboDojo benchmark (HKU MMLab-led) with a 32.26% average success / 39.26 score — ahead of GPT-6-Astra (22.48%), π0.5 (8.31%) and NVIDIA GR00T-N1.7 — with no extra training data [1].
2. Target Users & Pain Points
- Logistics / industrial automation — deployed at SF Express and China Post sorting centers across 5 provinces; robots sort up to
1,200 parcels/hour (85% of human efficiency) [4]. - General robotics developers — the open-source VPP2 model gives third parties a frontier-tier manipulation policy.
Pain solved: teaching robots dexterous manipulation normally requires huge teleop/RL data; the video-prediction approach learns from passive video, collapsing the data and compute barrier to generalist robot control [1].
3. Competitive Landscape
| Player | Approach | Vs. Robot Era |
|---|---|---|
| unitree | Low-cost humanoids + RL | Hardware-first, price leader; weaker open policy stack |
| galaxea / agibot | Humanoids + data flywheel | China peers; Robot Era's edge is open world-action models |
| physical-intelligence | π0.x foundation models | US frontier; VPP2 beat π0.5 on RoboDojo |
| figure-ai | Humanoids + Helix | US commercial focus; Robot Era is logistics-deployed in China |
Robot Era's differentiation is the open-source world-action model — VPP2 is a public benchmark leader, which no major Western or Chinese peer has matched on the same board [1].
4. Unique Observations
- Prediction-decoupling as the winning recipe. VPP2's three-stage design (event-level video pre-training → 8-second clip distillation → 0.9B action DiT) lets a 14B model beat far larger frontier systems without extra data or "agent" RSI tricks — evidence that the embodied-ai field is converging on video prediction as the core pre-training objective [1].
- A public, model-first benchmark moat. Unlike hardware-led peers, Robot Era wins mindshare through the model leaderboard — 58.5% zero-shot on 10 real ALOHA tasks (vs π0.5's 40%) is the number third parties will cite [1].
- Institutional lock-in. It is the only embodied-AI company Tsinghua directly holds equity in (spawned from 交叉信息研究院, founder recruited by Andrew Yao) — giving it preferential access to talent and state-backed capital that a purely commercial startup lacks [3][4].
5. Financials / Funding
- Jul 2025: ~RMB 500M A round.
- Nov 2025: ~RMB 1B A+ round, led by Geely Capital, with BAIC strategic investment and Beijing AI-industry / robotics funds [2].
- Mar 2026: ~RMB 1B strategic round; valuation surpassed RMB 10B (unicorn) [4].
- Apr 2026: >US$200M round led by SF Express, with Sequoia China and IDG [4].
- Jul 2026: ~RMB 1B Series B, led by state-owned Chengtong Fund with regional state funds [3].
- Cumulative: media reports put total raised near RMB 10B (~US$550M+); cumulative orders >RMB 500M, ~50% overseas [3][4].
6. People & Relationships
- Founder / CEO: Chen Jianyu (陈建宇) — Tsinghua assistant professor, UC Berkeley PhD under Masayoshi Tomizuka, recruited back to Tsinghua by Andrew Yao [3][4].
- Investors: Geely Capital, BAIC, SF Express, Sequoia China, IDG, Chengtong Fund, Alibaba, Legend Capital, Hillhouse.
- Partners / deployments: SF Express, China Post (logistics sorting).
- Peers: unitree, galaxea, agibot, physical-intelligence, figure-ai.