What the RACER-Knee Trial Actually Tested
RACER-Knee found no functional benefit from robotic-arm assistance in TKA. Both arms received a CT scan and a preoperative plan, which changes what the result means.
Key takeaways
RACER-Knee is the largest blinded randomised trial of robotic-arm assisted total knee arthroplasty, and its primary outcome was null: no difference in the Forgotten Joint Score at 12 months. Radiographic accuracy improved, cost went up, and the patient felt nothing. The detail that most summaries drop is in the authors' own limitations paragraph: both arms received a CT scan and a preoperative plan. The trial isolated the robotic arm, not the plan. That makes it a clean result about hardware and an open question about planning, and it would be dishonest to read it as evidence that planning works. A second 2026 trial, RASKAL, tested the plan's most important output under double-blind randomisation and also found nothing.
What the trial did
Parsons H, Metcalfe A, Griffin J, et al. Lancet 2026;408(10558):924-934, first published online 20 August 2026. doi:10.1016/S0140-6736(26)00986-4 · PMID 42624811 · ISRCTN27624068
n=339 (168 robotic, 171 conventional), 10 hospitals in Great Britain, 33 surgeons. Participants and outcome assessors were blinded, with sham pin incisions used to maintain masking. This is an unusually rigorous design for a surgical device trial, where blinding is normally the first thing to be abandoned.
The results
| Outcome | Robotic | Conventional | Result |
|---|---|---|---|
| Forgotten Joint Score, 12 months (primary) | 49.2 (SD 28.4) | 50.2 (SD 29.9) | Adjusted difference -1.5 (95% CI -7.5 to 4.5), p=0.62 |
| Deviation from planned HKA | 2.0° (1.8) | 2.8° (2.4) | p=0.0027, favours robotic |
| Deviation from planned tibial slope | 3.1° (2.2) | 4.4° (2.7) | p<0.0001, favours robotic |
| Cost at 12 months | More costly | Reference | Not cost-effective |
Note the direction of the primary result: the point estimate favours the conventional arm, and the confidence interval comfortably contains zero.
A caution on figures circulating in summaries and conference slides: the Lancet abstract states only that robotic surgery was more costly, without a number. Specific cost and operative-time deltas attributed to this trial are widely repeated but are not in the openly accessible record. We do not quote them, and neither should a pitch deck.
The sentence most summaries drop
From the trial's own limitations:
"CT-based planning was used but the effects were not evaluated in this trial, and the findings therefore do not establish whether CT-based planning itself provides benefit. A CT scan and a preoperative plan were generated for both groups to maintain participant masking... the trial was designed to evaluate the additional effect of robotic-arm assistance during surgery."
Both arms were planned. The single variable was the robotic arm during surgery. So the defensible reading is narrow and specific: adding a robotic arm on top of CT-based planning did not change how the knee felt at one year. It is not a verdict on planning, in either direction.
The part that cuts against planning
This is where an honest reading has to keep going, because there is a second 2026 trial that does test something closer to the plan.
RASKAL (MacDessi et al., Bone Joint J 2026, PMID 41944850) randomised the alignment target itself in a double-blind 2x2 factorial design, n=303. The KOOS-12 difference between functional and mechanical alignment was 0.3 points (95% CI -3.4 to 4.0, p=0.867).
The intervention demonstrably took: soft-tissue release rates differed by a factor of 5.5 between arms. So the plan really did change, the operation really was performed differently, and the patient still felt no difference. Anyone using RACER to argue "it's the plan, not the robot" has to answer RASKAL, and most presentations do not.
The pattern is older than either trial
| Evidence | Finding |
|---|---|
| Arshi A, et al. J Arthroplasty 2022;37:1562-1569 | Eight robotic TKA studies, n=1,529. One of eight reached statistical significance on any patient-reported outcome; zero of eight reached the minimal clinically important difference |
| Kim YH, Yoon SH, Park JW. Clin Orthop Relat Res 2020;478:266-275 | Randomised, n=1,406, mean 13-year follow-up. 15-year survivorship 98% vs 98%. All endpoints null |
| AOANJRR 2025 annual report | 100,081 robotic TKAs, adjusted hazard ratio for revision 1.00 (95% CI 0.93-1.07) |
The registry number carries a limitation worth stating: robotic follow-up cannot be read much past six years. At year seven there were 372 knees at risk; at year eight, eight. Nobody on either side of this argument has long-term robotic registry data.
What the health technology assessment said
NICE's early value assessment (17 April 2025) placed six robotic-assisted joint replacement platforms in conditional use during a three-year evidence-generation period, noting that patient-reported outcome gains sat below the threshold of clinical significance. The assessment covers hip and unicompartmental replacement as well as TKA.
One honest complication: the single peer-reviewed commentary on that same assessment reads it more favourably, describing potential cost-effectiveness particularly for hip (Clement ND, Haddad FS. Bone Joint J 2026;108-B(2):147-152, doi:10.1302/0301-620X.108B2.BJJ-2025-1438). Both readings exist and a fair summary includes both.
The second implication points at software companies, including ours. A planning tool entering the same category will meet the same evidence demand. The assessment asked for evidence generation, not marketing.
Citations you will see that do not hold
Working through this literature, several confident references turn out not to exist as described:
- There is no Song meta-analysis and no Liow meta-analysis comparing robotic with conventional TKA.
- The Cochrane review of robotic TKA is a protocol only (CD016286, registered February 2026). "A Cochrane review says" is not currently a true sentence here.
- Deckey, Held and Bhimani are retrospective cohorts, not randomised trials.
- The Glasgow trial (Blyth, Jones) is unicompartmental, not total, knee arthroplasty. Its results do not transfer.
- Operative-time penalties are usually quoted against manual instrumentation. Against computer-assisted navigation the difference has not been shown (Haslhofer 2024: 80.1 vs 78.3 minutes, p=0.451).
What nobody has tested
The comparison that would settle the planning question has not been run: three-dimensional CT planning versus no formal plan at all. Every trial in this literature plans both arms and then varies the execution tool or the target. That gap is real, and it is the reason we describe our own software as Research Use Only and frame engagements as evidence-building rather than as proven benefit.
It is also why we think the more tractable near-term question is not "does the plan improve outcome" but "is the measurement underneath the plan reproducible", which turns out to have an uncomfortable published answer.
Bottom line
RACER-Knee is a well-run trial with a null primary outcome, better radiographic accuracy, higher cost, and one crucial design detail: both arms were planned from CT, so the robotic arm was the only variable. Read narrowly it says that intraoperative robotics did not add patient-perceived benefit on top of a plan. Read alongside RASKAL it warns that the plan's own most important output has also failed a blinded randomised test. The honest position for anyone building planning software is that the burden of proof sits with us. For how the software-only and robotic approaches differ in what they actually ask of a hospital, see robotic versus software-only TKA planning and the 2026 comparison of orthopedic 3D planning software.
Reviewed by the Salnus biomedical engineering team.