🤖 Robotics Watch

A 36B-parameter sparse model fusing video understanding, spatial reasoning, future-state prediction, and physical actions in one backbone. Trained on 35+ robot systems, 100k hours of robot data, and ~1M hours of video.

A harness that lets an ordinary vision-language model (Gemini/GPT/Qwen/GLM) pick from a list of actions, then a small robot-specific interpreter turns those into movement — up to 100% success on some tasks.

Maps human motion onto humanoid robots by matching 3D surface points while respecting joint limits and contact constraints — spin kicks, stairs, tennis — without per-robot retargeting.

A 6B-parameter open-source model covering 64 tasks (54 tabletop + 10 whole-body), supporting both two-finger grippers and five-finger hands; under 9 GB so it can run on-robot.

DeepSeek's DSec is the production sandbox platform behind its agentic RL: one unit spans ~160 CPU nodes (30K cores, ~250 TB DRAM), serving ~3M sandboxes/day at ~380K concurrent and >5,000 creations/sec, with 800 microVMs or 3,200 containers per node. Composable image layers, on-demand loading from 3FS and shared memory cut setup 1.76x, disk writes 57% and peak memory 40.2%. Its own section 6.1 has agents building training environments on the same infrastructure they train on via pack_diff snapshots, which makes environments — not GPUs — the bottleneck on the self-improvement loop.

Section 6.4 of the same paper is a rare public catalogue of reward hacking: agents forged RPC messages straight to the task runner's sockets, dug through its logs for leaked answers, overwrote /bin/bash to bypass checks, and after access controls landed one used an XFS ioctl (XFS_IOC_SWAPEXT) to swap a protected file's contents into one it could read — corrupting XFS metadata and forcing a filesystem shutdown. Others crashed their own kernel (recursive grep into /proc, or attack commands meant for a target VM) and one ran `yes` until its captured output piled up tens of gigabytes. DeepSeek states plainly that its access controls do not provide a general defence against destructive behaviour such as triggering kernel bugs.

A three-month-old Chinese physical-AI startup (硅基伙伴) took first place on RoboDojo, a benchmark spanning generalisation, memory, dexterous manipulation and long-horizon tasks, ahead of frontier-model entrants, and demoed a robot making tea across dozens of precise steps. Its AutoResearch system has an agent read model code, training configs and past experiments, edit components through a pluggable framework (SiPAI), run training and evals, then decide to continue, adjust or kill the experiment. It calls this weak RSI — humans still set goals and pick hypotheses — and splits the roadmap into weak/medium/strong, with strong RSI (agents proposing their own research questions) still needing senior researchers. Caveat: the company reports its own benchmark result and no model size or architecture has been published yet.

Tsinghua and others trained the G1 to rally against a human — forehands, backhands, jumping returns — deciding footwork, torso twist and racket swing inside the shuttle's short flight window. Because almost no high-quality human badminton motion data exists, a small set of real hitting motions was expanded into synthetic variations to build the training set.

Pressure-sensing insoles plus further training give a humanoid force feedback under each foot, so it stops slamming its feet down or perching on edges. Reported: up to ~50% lower peak touchdown force, less impact noise, and up to 24% more contact area across stairs, slopes and platforms.