One-Line Summary

Three papers tighten the link between intent and execution: a single whole‑body humanoid controller, a speed‑controllable robot policy, and unified object‑plus‑material selection.

Research Papers

HANDOFF unifies humanoid whole-body control with distilled experts

HANDOFF is a single controller for humanoid robots that takes a compact, explicit command interface and turns it into coordinated whole‑body motion—so planners can express tasks without crafting dense kinematic or spatial references. ¹

The system learns by distilling three complementary specialists—whole‑body motion tracking with safety‑filtered data, locomotion, and fall recovery—into a Mixture of Experts (MoE) student using a context‑conditioned gating scheme and Kullback–Leibler (KL) divergence‑based distillation. ¹

On the Unitree G1 platform, HANDOFF matches state‑of‑the‑art velocity tracking and offers one of the largest robust manipulation workspaces reported, while remaining feasible on hardware. ¹

It also drives multiple natural‑language tasks via a Vision‑Language Model (VLM)‑powered agentic planner without task‑specific data or controller fine‑tuning—pointing to a cleaner “command space” between planning and whole‑body control that generalizes across skills. ¹

TempoVLA adds on-demand speed control to robot policies

TempoVLA lets one Vision‑Language‑Action (VLA) policy speed up on low‑risk transits and slow down for high‑risk contact, instead of being stuck at a single pace inherited from demonstrations. ²

It introduces Variable‑Speed Trajectory Augmentation (VSTA) to retime demos by merging or splitting actions to hit a target speed while preserving motion semantics, and conditions the policy on the requested speed; reported statistics show VSTA reaches the requested speed with negligible motion error, and it even boosts the default 1× performance through better data use. ²

In simulation and real tasks, TempoVLA achieves flexible bidirectional speed control; paired with a large multimodal model, it dynamically accelerates through easy phases and decelerates for delicate ones—complementing prior attempts like Key‑Value (KV) cache reuse or reinforcement learning (RL) that shift policies to another fixed speed. ²

MAOAM unifies object and material selection for image editing

MAOAM is a selection framework that produces pixel‑accurate masks from either text prompts or clicks, and it can select not just objects but also materials (for example, all the wood or glass in a scene) to enable practical re‑texturing and consistent edits. ³

It uses a Vision‑Language Model (VLM) with a segmentation head: the VLM interprets intent across entities, attributes, and spatial relations, while the head maps output tokens to a mask; to overcome scarce material‑selection data with text, the authors build a scalable pipeline combining real and synthetic images with material masks and VLM‑generated descriptions, then train with multi‑task objectives over click and text plus an auxiliary Visual Question Answering (VQA) task—yielding accurate, coherent selections and emergent gains when combining text and clicks at inference. ³

Why It Matters

Together, these papers push intent‑driven control closer to practice: a compact command interface lets planners speak the robot’s language (HANDOFF), explicit speed control keeps execution safe and efficient (TempoVLA), and richer selection semantics make visual tools more faithful to what users mean (MAOAM). ¹

For non‑robotics teams, the pattern is the same: add a thin, controllable layer between high‑level reasoning and low‑level actuation, then train it with the right data augmentations or expert distillation—an approach that can reduce brittle hand‑tuning and make systems safer to deploy. ²