Leaderdrive's first three quarters revenue up 45%, humanoid reducer orders reach 400,000 units
Leaderdrive's revenue for the first three quarters of 2026 was 591 million yuan, up 45.30% year on year, with net profit attributable to shareholders of 122 million yuan, up 30.33% year on year; humanoid-related reducer orders rose from about 100,000 units the previous year to 400,000 units. The company's annual harmonic reducer production capacity has reached 1 million units, with a year-end target of 1.5 million units, but its gross margin in the first half was 32.26%, down 2.51 percentage points year on year, as new production line commissioning and small-batch production of multiple models are weighing on profitability. The article also compares the catch-up pace of domestic manufacturers such as Zhongda Leader, Huandong Technology and Laifu Harmonic.
Highlights
GE-Sim 2.0 and π0.5 closed-loop simulation tests: why reward hacking occurs in VLA training
The Xbotics embodied intelligence lab used the AgiBot GE-Sim 2.0 world model and the OpenPI π0.5 policy to build a closed-loop simulation pipeline, measuring about 4.33 frames per second for world model generation, about 3.20 frames per second for the pure policy closed loop, and about 0.88 frames per second after adding MiMo online scoring.
Xbotics具身智能实验室 · 微信公众号Source ↗Peking University-Lingchu Joint Lab's RoboFFT accepted by CoRL2026, using reinforcement learning post-training to generate generative robot policies
RoboFFT, proposed by the Peking University-Lingchu Joint Lab, has been accepted by CoRL2026. It is a reinforcement learning post-training method for generative robot policies that reuses the forward noising process from pretraining of diffusion models and flow models to construct approximate probability ratios for PPO-style clipped updates.
Lingchu Intelligence and others propose FAFM: a continuous action generation method for heterogeneous-frequency data, accepted by NeurIPS 2026
Lingchu Intelligence, together with Beihang University, Peking University and the Chinese University of Hong Kong, proposed the frequency-aware flow matching method FAFM, whose paper was accepted to the NeurIPS 2026 main conference. The method uses the discrete cosine transform (DCT-II) to represent actions as frequency-domain coefficients corresponding to real time, performs flow matching in coefficient space and constrains the first-order time derivative, requires no additional learnable parameters, and can be used independently or integrated into VLA models such as π₀ and π₀.₅.
Industry dynamics
Leaderdrive's first three quarters revenue up 45%, humanoid reducer orders reach 400,000 units
Leaderdrive's revenue for the first three quarters of 2026 was 591 million yuan, up 45.30% year on year, with net profit attributable to shareholders of 122 million yuan, up 30.33% year on year; humanoid-related reducer orders rose from about 100,000 units the previous year to 400,000 units. The company's annual harmonic reducer production capacity has reached 1 million units, with a year-end target of 1.5 million units, but its gross margin in the first half was 32.26%, down 2.51 percentage points year on year, as new production line commissioning and small-batch production of multiple models are weighing on profitability. The article also compares the catch-up pace of domestic manufacturers such as Zhongda Leader, Huandong Technology and Laifu Harmonic.
人形大讲堂 · 微信公众号Source ↗
Paper research
GE-Sim 2.0 and π0.5 closed-loop simulation tests: why reward hacking occurs in VLA training
The Xbotics embodied intelligence lab used the AgiBot GE-Sim 2.0 world model and the OpenPI π0.5 policy to build a closed-loop simulation pipeline, measuring about 4.33 frames per second for world model generation, about 3.20 frames per second for the pure policy closed loop, and about 0.88 frames per second after adding MiMo online scoring.
Xbotics具身智能实验室 · 微信公众号Source ↗Peking University-Lingchu Joint Lab's RoboFFT accepted by CoRL2026, using reinforcement learning post-training to generate generative robot policies
RoboFFT, proposed by the Peking University-Lingchu Joint Lab, has been accepted by CoRL2026. It is a reinforcement learning post-training method for generative robot policies that reuses the forward noising process from pretraining of diffusion models and flow models to construct approximate probability ratios for PPO-style clipped updates.
Lingchu Intelligence and others propose FAFM: a continuous action generation method for heterogeneous-frequency data, accepted by NeurIPS 2026
Lingchu Intelligence, together with Beihang University, Peking University and the Chinese University of Hong Kong, proposed the frequency-aware flow matching method FAFM, whose paper was accepted to the NeurIPS 2026 main conference. The method uses the discrete cosine transform (DCT-II) to represent actions as frequency-domain coefficients corresponding to real time, performs flow matching in coefficient space and constrains the first-order time derivative, requires no additional learnable parameters, and can be used independently or integrated into VLA models such as π₀ and π₀.₅.