HALO — Perceptive Humanoid Motion from Simulation was released and achieved 91% success in 360 physical trials. The system runs fully onboard and the paper, code, and project page are live.
Key points
Trained entirely on synthetic data from 1,000 procedurally generated homes.
Zero-shot deployment in 6 unseen real environments with 91% success across 360 physical trials.
Sources
XMimic trains one policy on 40 LAFAN1 clips in 54 seconds
XMimic trains one policy on all 40 LAFAN1 clips in 54 seconds using 4 H100 GPUs. It is 3.5× faster than BlinkMimic and 85× faster than SuperMimic.
Key points
Trains one policy on all 40 LAFAN1 clips in 54 seconds on 4 H100s.
3.5× faster than BlinkMimic and 85× faster than SuperMimic.
Sources
MedHEB benchmark released with 107 tasks and MedEmb baseline
MedHEB introduces a unified benchmark covering 107 retrieval tasks from 54 datasets across image, volume, and text modalities. It evaluates embedding models on modality-bridging and region-level retrieval, revealing uneven performance. The authors also release MedEmb, an MLLM-based reference embedding model trained on the MedHEB ID split.
Key points
Benchmark includes 107 retrieval tasks from 54 datasets across three evidence axes.
Provides MedEmb, a reproducible baseline model for 2D-3D-text medical retrieval.
Sources
s³ multi-agent RL paper shows 166% gain, 3x scaling
s³ multi-agent RL lets LLM swarms self-organize and beat best-of-N sampling, improving token efficiency and success rate by up to 166%. The approach scales to swarms up to three times larger than trained.
Key points
Asynchronous multi-agent training with no prescribed roles yields emergent coordination and communication protocols.
Improves token efficiency and success rate over compute-balanced baselines by as much as 166% and generalizes to up to 3× larger swarms.
Sources
OpenAI March 2028 RSI target, Anthropic 2027 RSI target
OpenAI claims it will achieve recursive self-improvement by March 2028. Anthropic employees say they aim for the same capability by 2027.
Key points
OpenAI on track to achieve RSI by March 2028
Anthropic employees on track to achieve RSI by 2027
Sources
Ctrl-F Force as an Interface for Humanoid Loco-Manipulation announced
Ctrl-F is introduced as a force-based interface for humanoid locomotion and manipulation. It addresses the underspecified interaction of reference motion by adding force.
Key points
Ctrl-F: Force as an Interface for Humanoid Loco-Manipulation.
Reference motion tells a humanoid where to move, but leaves the interaction underspecified.
Sources
- Original post
- forcectrl.github.io - Linked in the post
- arxiv.org - Linked in the post
INTERACT Learning Interaction-Aware World Dynamics paper released
INTERACT: Learning Interaction-Aware World Dynamics from Driving Videos is released as arXiv:2611.02526, a three-stage object-centric model. It learns recurrent slots from unlabeled videos, infers latent ego actions, and distills predictions into a trajectory planner, achieving state-of-the-art results on NAVSIM with better collision metrics.
Key points
Three-stage framework: slot-based scene representation, latent ego action inference, trajectory planner distillation.
State-of-the-art performance on NAVSIM, with improvements in safety-critical collision metrics.
Sources
SeMo3D unifies 3D generation, editing, and completion
SeMo3D introduces a unified approach for 3D generation, editing, and completion using a Selective Modeling paradigm. The paper proposes region-selective denoising training and an Omni3DiT diffusion transformer that handles image, text, and point guidance. Experiments on three benchmarks show SeMo3D surpasses state-of-the-art methods on all tasks.
Key points
Authors: Xiaokun Sun, Jian Yang, Zhenyu Zhang.
ArXiv ID: 2611.02568 (cs.CV).
Introduces Selective Modeling to synthesize a designated region while preserving the rest.
Proposes Omni3DiT, a multi-condition 3D diffusion transformer.
Sources
- Original post
- SeMo3D: Unifying 3D Generation, Editing, and Completion via Selective Modeling - Linked page
- xiaokunsun.github.io - Linked in the post
