Robotics Briefing /
From individual skills to complete robot tasks.
A robot can learn to sit, grasp or push and still struggle to combine those skills into a useful job. Today's research makes that gap visible: simulated homes teach a humanoid to move through real spaces, touch signals improve contact, and broader practice helps a manipulation model generalize. A distributed physical benchmark adds another lesson: the strongest model depends on what is tested and how it is trained.
Your quick takeaways
- HALO reports 91% success across 360 physical single-skill trials; its smaller test of chained tasks averaged 70%.
- Combining tactile cues raised success from 52 to 71 out of 90 attempts in a controlled three-task comparison.
- Multi-task reinforcement learning improved transfer to new objects in simulation, but did not establish the same gains on physical robots.
Papers
HALO: synthetic homes, physical movement
Siyuan Hu and colleagues at Apple · Preprint first submitted November 4 · Simulation and physical tests
HALO creates paired camera views, language instructions and motion examples inside 1,000 generated homes. A motion generator turns instructions and images into a movement plan; a separate controller uses depth and body-state feedback to execute it. A higher-level model breaks longer requests into subtasks.
On a modified Unitree G1, the authors report 91% average success across 360 trials in six previously unseen physical environments. The six skills include sitting, passing through partly open doors, crawling beneath obstacles and pushing boxes. Three chained tasks averaged 70% in 30 trials in one physical environment.
The difference matters: a successful crouch sometimes left the robot facing away from its next target. This is evidence of transfer from simulation, with a limited skill repertoire. Humans still specify new tasks and motion keyframes, and missing visual contact can disrupt interaction.
What touch contributes after contact begins
Xuhao Qin and colleagues · First submitted November 4; accepted to IEEE ROBIO 2026 · Physical tests
Camera-based tactile sensors see a soft fingertip deform. This study separates that stream into appearance, surface geometry and marker flow—the movement of embedded dots as the fingertip stretches sideways. The researchers compare those cues using the same demonstrations and policy backbone.
In the main comparison, combined touch signals achieved 71 successes in 90 attempts, versus 52 with vision alone. Marker flow was the strongest individual cue overall, at 67 out of 90. Tasks covered test-tube reorientation and insertion, blackboard wiping, and peg insertion.
Touch therefore supplies useful information that external cameras miss, particularly while maintaining contact. But the full combination beat each task's best individual cue by only one successful attempt. The complete study includes 720 trials across eight configurations; it uses one sensor, one gripper design and three tabletop tasks. Different cue-processing networks and the absence of direct force measurements also limit causal conclusions.
Broader practice can improve more than the practiced tasks
Michal Korniak and colleagues · Preprint first submitted November 4 · Simulation only
Fine-tuning a robot on one task can make it worse at others. This study asks whether reinforcement learning—improvement from attempted actions and their outcomes—can instead strengthen a generalist policy across many tasks.
Using pi0.5 in MolmoSpaces and LIBERO, the researchers compare learning from successful attempts with methods that train a critic: a model estimating how promising an action is. As training expands across objects and scenes, critic-based methods shift from forgetting toward generalization. In MolmoSpaces, gains over the starting policy reach 58 percentage points on unseen scenes and 39 points on unseen objects.
Those are best reported gains across the studied settings, not universal improvements. The strongest evidence concerns simulated grasp-and-lift tasks. More varied LIBERO tasks remained difficult to generalize across, and the starting policy needed some competence. The result suggests that both practice diversity and the learning rule matter; physical deployment remains a separate test.
RoboColosseum: rankings depend on the test
Tan Sang Nguyen and colleagues · Catch-up: preprint first submitted November 2 · Distributed physical evaluation
RoboColosseum compares robot policies on locally operated hardware under shared protocols. Its Open Track collected 5,699 blinded pairwise comparisons at 124 sites. A separate Fine-tune Track supplies each model with the same 100 demonstrations per task before testing adaptation.
Rankings changed substantially between tracks. On Franka, pi0.5 moved from fifth among the evaluated vision-language-action policies in the Open Track to first after fine-tuning. That does not isolate the effect of training alone: the tracks also differ in tasks, starting checkpoints and ranking metrics.
The study makes partial progress visible alongside complete success. None of the evaluated policies completed any of the five G1 humanoid tasks in the Fine-tune Track, despite making differing amounts of progress. Comparing intermediate steps can reveal where a system fails, but reaching those steps should not be reported as completing the job.
News
Liber-0 tests the bridge from human experience to robot actions
LiberAI · November 4 press announcement · Company-run physical evaluations
LiberAI describes Liber-0 as a world action model that jointly predicts future images and actions. Its training combines broad human video with measured hand movements from sensor gloves and robot recordings. The company reports more than 2.2 million hours of human experience; this does not mean 2.2 million hours of robot demonstrations.
On 60 tasks proposed by GPT-6, with ten trials each, LiberAI reports 36% overall success without robot demonstrations collected specifically for those tasks. Results range from 63.5% for pick-and-place to 10.5% for challenging multi-step tasks. These are company results, not an independent benchmark.
Pairing video with measured movement could help address scarce robot training data, but the multi-step result leaves considerable room for improvement. The featured milk-tea demonstration uses application-specific adaptation and is separate from the zero-shot suite. The technical report provides demonstrations and methods, but no downloadable model weights.
Highlights
One code release, one pending
Resources accompanying the November 4 research
The generalist-improvement repository contains a minimal implementation of the three compared learning approaches, with training scripts and environment setup for pi0.5 on LIBERO-90 and MolmoSpaces. It makes the algorithmic comparison inspectable, although reproducing it requires substantial compute: the README specifies a four-GPU node with roughly 100 GB of memory per GPU.
HALO's project offers an annotated demonstration gallery and a view of its autonomy stack. Its Code button is currently disabled and marked coming soon. The videos are available; a complete software release should not yet be assumed.
Read this first
Read RoboColosseum's section on performance consistency. Its explanation of why changed rankings cannot be attributed to fine-tuning alone is a useful guide to comparing robot leaderboards.
Read the paper